{
  "person": {
    "name": "Ian Pinto",
    "url": "https://ainpm.in/about/",
    "github": "https://github.com/opnsrcntrbtr",
    "huggingface": "https://huggingface.co/opnsrcntrbtrian",
    "credly": "https://www.credly.com/users/ian-pinto/badges"
  },
  "generated": "2026-11-16",
  "method": "Skills are written in employer language, aligned to O*NET work activities and NACE competencies, and linked to public evidence. Credentials link to the issuer's verification page; self-uploaded records are labelled.",
  "skills": [
    {
      "id": "ai-evaluation",
      "name": "AI evaluation design",
      "claim": "Turn a product requirement into a graded test set with metric floors that block a release.",
      "proof": "260-query adjudicated golden set; abstention accuracy 0.981 and recall@10 0.943 on SEBI Circular RAG.",
      "onet": [
        "Evaluating Information to Determine Compliance with Standards",
        "Analyzing Data or Information"
      ],
      "nace": "Critical Thinking",
      "evidence": [
        {
          "label": "SEBI RAG evaluation log",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/status.md",
          "kind": "Eval results"
        },
        {
          "label": "Evals are the new PRD",
          "href": "https://ainpm.in/articles/evals-are-the-new-prd/",
          "kind": "Publication"
        },
        {
          "label": "Lecture Lens local eval harness",
          "href": "https://ainpm.in/work/lecture-lens/",
          "kind": "Case study"
        },
        {
          "label": "CodeSearchNet retrieval report: baselines, ablation, significance tests",
          "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md",
          "kind": "Eval results"
        },
        {
          "label": "Merged fix with 5 regression tests (graphify #576)",
          "href": "https://github.com/Graphify-Labs/graphify/pull/576",
          "kind": "Contribution"
        },
        {
          "label": "Databricks Generative AI Fundamentals",
          "href": "https://www.credential.net/46362a83-a223-47bb-a783-1c522b65a1fd",
          "kind": "Credential"
        },
        {
          "label": "Judge calibration on a blind holdout (Lecture Lens origin)",
          "href": "https://ainpm.in/work/lecture-lens/origin/#proved",
          "kind": "Case study"
        },
        {
          "label": "Builder's note: is your LLM judge good enough?",
          "href": "https://ainpm.in/articles/is-your-llm-judge-good-enough/",
          "kind": "Publication"
        }
      ]
    },
    {
      "id": "rag-product",
      "name": "Retrieval (RAG) product design",
      "claim": "Design retrieval products that cite sources, track which documents are still in force and refuse when evidence is missing.",
      "proof": "1,490 SEBI circulars indexed with supersession-aware citations and as-of-date answers.",
      "onet": [
        "Getting Information",
        "Processing Information"
      ],
      "nace": "Technology",
      "evidence": [
        {
          "label": "Live demo on Hugging Face",
          "href": "https://huggingface.co/spaces/opnsrcntrbtrian/sebi-circular-rag-demo",
          "kind": "Live demo"
        },
        {
          "label": "sebi-circular-rag repository",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag",
          "kind": "Open code"
        },
        {
          "label": "SEBI circulars dataset (CC-BY-4.0)",
          "href": "https://huggingface.co/datasets/opnsrcntrbtrian/sebi-circulars",
          "kind": "Open code"
        },
        {
          "label": "PM case study",
          "href": "https://ainpm.in/sebi-rag/",
          "kind": "Case study"
        },
        {
          "label": "CodeSearchNet Find & Explain live demo",
          "href": "https://huggingface.co/spaces/opnsrcntrbtrian/csne-find-and-explain",
          "kind": "Live demo"
        },
        {
          "label": "Fine-tuned retrieval model (MiniLM)",
          "href": "https://huggingface.co/opnsrcntrbtrian/csne-minilm-retrieval-full-1ep",
          "kind": "Open code"
        },
        {
          "label": "Google Cloud BigQuery skill badges",
          "href": "https://www.credly.com/users/ian-pinto/badges",
          "kind": "Credential"
        }
      ]
    },
    {
      "id": "trust-compliance",
      "name": "Trust and compliance by design (India)",
      "claim": "Translate DPDP, RBI FREE-AI, SEBI and MeitY obligations into product controls in the first sprint.",
      "proof": "Regulation radar with dated obligations; abstention and citation controls shipped in a regulated-domain product.",
      "onet": [
        "Evaluating Information to Determine Compliance with Standards",
        "Monitoring Processes, Materials, or Surroundings"
      ],
      "nace": "Professionalism",
      "evidence": [
        {
          "label": "India AI regulation for PMs",
          "href": "https://ainpm.in/articles/india-ai-regulation-for-product-managers/",
          "kind": "Publication"
        },
        {
          "label": "Regulation radar",
          "href": "https://ainpm.in/guide/#radar",
          "kind": "Publication"
        },
        {
          "label": "SEBI RAG trust principles",
          "href": "https://ainpm.in/sebi-rag/",
          "kind": "Case study"
        },
        {
          "label": "SEBI RAG live demo with supersession-aware citations",
          "href": "https://huggingface.co/spaces/opnsrcntrbtrian/sebi-circular-rag-demo",
          "kind": "Live demo"
        },
        {
          "label": "SEBI RAG evaluation gate (abstention accuracy 0.981)",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/status.md",
          "kind": "Eval results"
        },
        {
          "label": "Respecting protected video while capturing lectures",
          "href": "https://ainpm.in/articles/respecting-protected-video-while-capturing-lectures/",
          "kind": "Publication"
        }
      ]
    },
    {
      "id": "privacy-security",
      "name": "Privacy and security architecture",
      "claim": "Specify zero-knowledge storage and passkey-based key handling, with a written threat model.",
      "proof": "Published npm library, a DOI-registered preprint and a PWA with 4,139 passing tests (6 Jul 2026).",
      "onet": [
        "Working with Computers",
        "Thinking Creatively"
      ],
      "nace": "Technology",
      "evidence": [
        {
          "label": "webauthn-prf-zktv on npm",
          "href": "https://www.npmjs.com/package/webauthn-prf-zktv",
          "kind": "Package"
        },
        {
          "label": "Preprint (DOI)",
          "href": "https://doi.org/10.6084/m9.figshare.32915411.v1",
          "kind": "Publication"
        },
        {
          "label": "TrustVault repository",
          "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA",
          "kind": "Open code"
        },
        {
          "label": "TrustVault case study",
          "href": "https://ainpm.in/work/trustvault/",
          "kind": "Case study"
        },
        {
          "label": "password-suite on npm",
          "href": "https://www.npmjs.com/package/password-suite",
          "kind": "Package"
        }
      ]
    },
    {
      "id": "local-ai",
      "name": "On-device and local AI",
      "claim": "Ship AI features that run on the user’s device so personal data never leaves it.",
      "proof": "Lecture Lens runs transcription, notes and flashcards entirely on a Mac; TrustVault explains password risk with on-device models.",
      "onet": [
        "Working with Computers",
        "Updating and Using Relevant Knowledge"
      ],
      "nace": "Technology",
      "evidence": [
        {
          "label": "Lecture Lens docs",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/",
          "kind": "Open code"
        },
        {
          "label": "Lecture Lens case study",
          "href": "https://ainpm.in/work/lecture-lens/",
          "kind": "Case study"
        },
        {
          "label": "Lecture Lens repository",
          "href": "https://github.com/opnsrcntrbtr/lecture-lens",
          "kind": "Open code"
        },
        {
          "label": "MLX AutoResearch agent (local MLX experiment loop)",
          "href": "https://github.com/opnsrcntrbtr/mlx-autoresearch-agent",
          "kind": "Open code"
        },
        {
          "label": "Free learning path, module 3: models and cost per task",
          "href": "https://ainpm.in/learn/#m3",
          "kind": "Publication"
        },
        {
          "label": "Lecture Lens Studio: the model runs inside Chrome",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/studio/",
          "kind": "Live demo"
        },
        {
          "label": "Lecture Lens architecture: local-first pipeline and decisions",
          "href": "https://ainpm.in/work/lecture-lens/architecture/",
          "kind": "Case study"
        }
      ]
    },
    {
      "id": "product-strategy",
      "name": "AI product strategy and MVP scoping",
      "claim": "Frame the user job, pick the smallest product that proves value and decide what not to build.",
      "proof": "Four open-source products taken from idea to a public demo, package or staging app.",
      "onet": [
        "Developing Objectives and Strategies",
        "Making Decisions and Solving Problems"
      ],
      "nace": "Leadership",
      "evidence": [
        {
          "label": "Portfolio",
          "href": "https://ainpm.in/work/",
          "kind": "Case study"
        },
        {
          "label": "Eight-step playbook",
          "href": "https://ainpm.in/guide/#playbook",
          "kind": "Publication"
        },
        {
          "label": "What is AI-native PM?",
          "href": "https://ainpm.in/articles/what-is-ai-native-product-management/",
          "kind": "Publication"
        },
        {
          "label": "ISB Product Management certificate",
          "href": "https://www.credential.net/cc9d4612-76c4-47b7-8c95-d6b81ebe53fe",
          "kind": "Credential"
        },
        {
          "label": "AI product strategy framework for Indian teams",
          "href": "https://ainpm.in/articles/ai-product-strategy-framework-india/",
          "kind": "Publication"
        },
        {
          "label": "Lecture Lens product bets: measures, results and decisions",
          "href": "https://ainpm.in/work/lecture-lens/#bets",
          "kind": "Case study"
        }
      ]
    },
    {
      "id": "program-delivery",
      "name": "Technical program delivery",
      "claim": "Run cross-team delivery of cloud-native platforms with clear scope, risk and release control.",
      "proof": "Day-to-day work on mutual fund AMC products and platforms, backed by PMP (verifiable on Credly) and CSM (certificate on request).",
      "onet": [
        "Coordinating the Work and Activities of Others",
        "Scheduling Work and Activities"
      ],
      "nace": "Teamwork",
      "evidence": [
        {
          "label": "All credentials, with verification links",
          "href": "#credentials",
          "kind": "Credential"
        },
        {
          "label": "About Ian Pinto",
          "href": "https://ainpm.in/about/",
          "kind": "Case study"
        },
        {
          "label": "PMP on Credly (valid to Aug 2027)",
          "href": "https://www.credly.com/badges/af25e5e0-fabb-42a7-8083-4e4f1b9d9d51/public_url",
          "kind": "Credential"
        },
        {
          "label": "Merged feature in openselfservice (#365)",
          "href": "https://github.com/o2sdev/openselfservice/pull/365",
          "kind": "Contribution"
        },
        {
          "label": "express-swagger-auto on npm",
          "href": "https://www.npmjs.com/package/express-swagger-auto",
          "kind": "Package"
        }
      ]
    },
    {
      "id": "communicating-quality",
      "name": "Communicating AI quality honestly",
      "claim": "Explain what an AI product does well and badly, in numbers, to non-specialists.",
      "proof": "Case studies publish weak metrics too, such as citation precision 0.194, alongside rejected experiments.",
      "onet": [
        "Communicating with People Outside the Organization",
        "Interpreting the Meaning of Information for Others"
      ],
      "nace": "Communication",
      "evidence": [
        {
          "label": "Articles",
          "href": "https://ainpm.in/articles/",
          "kind": "Publication"
        },
        {
          "label": "Glossary",
          "href": "https://ainpm.in/glossary/",
          "kind": "Publication"
        },
        {
          "label": "SEBI RAG results, including weak ones",
          "href": "https://ainpm.in/sebi-rag/",
          "kind": "Eval results"
        },
        {
          "label": "CodeSearchNet case study: limits stated up front",
          "href": "https://ainpm.in/work/codesearchnet/",
          "kind": "Case study"
        },
        {
          "label": "graphify #576: a fix explained for beginners and experts",
          "href": "https://github.com/Graphify-Labs/graphify/pull/576",
          "kind": "Contribution"
        }
      ]
    }
  ],
  "credentials": [
    {
      "name": "Project Management Professional (PMP)",
      "issuer": "Project Management Institute",
      "category": "certification",
      "verifyUrl": "https://www.credly.com/badges/af25e5e0-fabb-42a7-8083-4e4f1b9d9d51/public_url",
      "verification": "issuer",
      "platform": "Credly",
      "issued": "2015-08-21",
      "expires": "2027-08-21",
      "skills": [
        "program-delivery",
        "communicating-quality"
      ]
    },
    {
      "name": "Generative AI Overview for Project Managers",
      "issuer": "Project Management Institute",
      "category": "badge",
      "verifyUrl": "https://www.credly.com/users/ian-pinto/badges",
      "verification": "issuer",
      "platform": "Credly",
      "issued": "2023-09-28",
      "skills": [
        "program-delivery",
        "product-strategy"
      ]
    },
    {
      "name": "Product Management (Online Executive Certificate Programme)",
      "issuer": "Indian School of Business (ISB) Executive Education",
      "category": "certificate",
      "verifyUrl": "https://www.credential.net/cc9d4612-76c4-47b7-8c95-d6b81ebe53fe",
      "verification": "issuer",
      "platform": "Accredible",
      "issued": "2022-04-01",
      "note": "Programme 29 Nov 2021 to 14 Mar 2022, delivered with Emeritus",
      "skills": [
        "product-strategy"
      ]
    },
    {
      "name": "Academy Accreditation: Generative AI Fundamentals",
      "issuer": "Databricks Academy",
      "category": "badge",
      "verifyUrl": "https://www.credential.net/46362a83-a223-47bb-a783-1c522b65a1fd",
      "verification": "issuer",
      "platform": "Accredible",
      "issued": "2023-09-30",
      "skills": [
        "rag-product",
        "ai-evaluation"
      ]
    },
    {
      "name": "Build a Data Warehouse with BigQuery (skill badge)",
      "issuer": "Google Cloud",
      "category": "badge",
      "verifyUrl": "https://www.credly.com/users/ian-pinto/badges",
      "verification": "issuer",
      "platform": "Credly",
      "issued": "2024-11-19",
      "skills": [
        "rag-product"
      ]
    },
    {
      "name": "Streaming Analytics into BigQuery (skill badge)",
      "issuer": "Google Cloud",
      "category": "badge",
      "verifyUrl": "https://www.credly.com/users/ian-pinto/badges",
      "verification": "issuer",
      "platform": "Credly",
      "issued": "2024-12-01",
      "skills": [
        "rag-product"
      ]
    },
    {
      "name": "Data Science programme, Course 1",
      "issuer": "upGrad",
      "category": "certificate",
      "verifyUrl": "https://www.credential.net/e9075bc5-d4fb-4546-8933-d0247e45e1b6",
      "verification": "issuer",
      "platform": "Accredible",
      "issued": "2020-08-18",
      "note": "Issued as \"DS C18 Course 1 LOA\"",
      "skills": [
        "ai-evaluation"
      ]
    },
    {
      "name": "AI Product Management",
      "issuer": "IIT Kharagpur with upGrad",
      "category": "certificate",
      "note": "Certificate available on request",
      "skills": [
        "product-strategy",
        "ai-evaluation"
      ]
    },
    {
      "name": "Certified ScrumMaster (CSM)",
      "issuer": "Scrum Alliance",
      "category": "certification",
      "note": "Certificate available on request",
      "skills": [
        "program-delivery"
      ]
    },
    {
      "name": "AWS Cloud Practitioner Essentials",
      "issuer": "Amazon Web Services",
      "category": "certificate",
      "verifyUrl": "https://www.credential.net/93e91ee4-acb1-4b14-b7ac-a452b086712c?username=ianpinto",
      "verification": "self-uploaded",
      "platform": "Accredible",
      "note": "Course completion, ID E-N0539V",
      "skills": [
        "privacy-security"
      ]
    },
    {
      "name": "MBA, Information Technology and Systems Management",
      "issuer": "NMIMS Global Access School for Continuing Education",
      "category": "degree",
      "verifyUrl": "https://www.credential.net/1d9b1ed0-221a-402e-80d2-42848a14d003?username=ianpinto",
      "verification": "self-uploaded",
      "platform": "Accredible",
      "note": "Semester transcripts uploaded by Ian",
      "skills": [
        "product-strategy"
      ]
    },
    {
      "name": "BE (Information Technology)",
      "category": "degree",
      "note": "Certificate available on request",
      "skills": [
        "privacy-security"
      ]
    }
  ],
  "projects": [
    {
      "name": "SEBI Circular RAG",
      "url": "https://ainpm.in/sebi-rag/",
      "summary": "Open-source retrieval-augmented assistant for Indian securities regulation with supersession-aware citations, as-of-date answers and a 260-query evaluation gate.",
      "status": "Live demo",
      "skills": [
        "ai-evaluation",
        "rag-product",
        "trust-compliance",
        "communicating-quality"
      ],
      "lastActivity": "2026-09-16",
      "activityCheckedOn": "2026-11-05",
      "links": [
        {
          "label": "Live demo",
          "href": "https://huggingface.co/spaces/opnsrcntrbtrian/sebi-circular-rag-demo"
        },
        {
          "label": "GitHub",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag"
        }
      ],
      "results": {
        "title": "Quality on 260 adjudicated questions",
        "asOf": "Full evaluation 15 Aug 2026 · gate floors derived 2 Sep 2026",
        "note": "Tick = release floor. Every metric clears its floor; citation precision is the known weak spot.",
        "source": {
          "label": "Evaluation log",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/status.md"
        },
        "metrics": [
          {
            "kind": "rate",
            "label": "Abstention accuracy",
            "value": 0.981,
            "floor": 0.9373
          },
          {
            "kind": "rate",
            "label": "Recall@10",
            "value": 0.943,
            "floor": 0.8397
          },
          {
            "kind": "rate",
            "label": "Context recall",
            "value": 0.916,
            "floor": 0.8192
          },
          {
            "kind": "rate",
            "label": "Citation recall",
            "value": 0.881,
            "floor": 0.7347
          },
          {
            "kind": "rate",
            "label": "nDCG@10",
            "value": 0.697,
            "floor": 0.5934
          },
          {
            "kind": "rate",
            "label": "Citation precision",
            "value": 0.194,
            "floor": 0.1466,
            "weak": true
          }
        ]
      }
    },
    {
      "name": "TrustVault PWA",
      "url": "https://ainpm.in/work/trustvault/",
      "summary": "React 19 progressive web app that encrypts every credential on the device, unlocks with WebAuthn PRF biometrics, and explains password risk with local AI models.",
      "status": "Staging",
      "skills": [
        "privacy-security",
        "local-ai",
        "program-delivery"
      ],
      "lastActivity": "2026-08-02",
      "activityCheckedOn": "2026-11-05",
      "links": [
        {
          "label": "Staging app",
          "href": "https://trust-vault-pwa.vercel.app"
        },
        {
          "label": "GitHub",
          "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA"
        }
      ],
      "results": {
        "title": "Engineering evidence",
        "asOf": "Project status report, 6 Jul 2026",
        "note": "Self-reported by the project; no independent audit yet.",
        "source": {
          "label": "Project status",
          "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
        },
        "metrics": [
          {
            "kind": "ratio",
            "label": "Automated tests passing",
            "num": 4139,
            "den": 4142
          },
          {
            "kind": "stat",
            "label": "Lighthouse PWA score",
            "value": "100"
          },
          {
            "kind": "stat",
            "label": "Encryption on device",
            "value": "AES-256"
          },
          {
            "kind": "stat",
            "label": "scrypt cost for key wrap",
            "value": "N=2¹⁷"
          }
        ]
      }
    },
    {
      "name": "webauthn-prf-zktv + preprint",
      "url": "https://ainpm.in/work/webauthn-prf-zktv/",
      "summary": "MIT-licensed TypeScript library that wraps a vault key under a WebAuthn PRF output or a scrypt password, with a threat model, interop vectors and a CC BY 4.0 preprint.",
      "status": "npm v0.2.0",
      "skills": [
        "privacy-security",
        "communicating-quality"
      ],
      "lastActivity": "2026-07-07",
      "activityCheckedOn": "2026-11-05",
      "links": [
        {
          "label": "npm",
          "href": "https://www.npmjs.com/package/webauthn-prf-zktv"
        },
        {
          "label": "Preprint (DOI)",
          "href": "https://doi.org/10.6084/m9.figshare.32915411.v1"
        },
        {
          "label": "GitHub",
          "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv"
        }
      ],
      "results": {
        "title": "Library at a glance",
        "asOf": "npm v0.2.0, 5 Jul 2026 · compatibility per README",
        "source": {
          "label": "README",
          "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv"
        },
        "metrics": [
          {
            "kind": "ratio",
            "label": "Platforms with WebAuthn PRF",
            "num": 5,
            "den": 6
          },
          {
            "kind": "stat",
            "label": "Version on npm",
            "value": "v0.2.0"
          },
          {
            "kind": "stat",
            "label": "Runtime dependency",
            "value": "1"
          },
          {
            "kind": "stat",
            "label": "Wrap schemes",
            "value": "2"
          }
        ]
      }
    },
    {
      "name": "Lecture Lens",
      "url": "https://ainpm.in/work/lecture-lens/",
      "summary": "Native macOS app, local-model toolkit and an in-browser Studio for working professionals in online courses, with a local evaluation harness and nothing uploaded.",
      "status": "Early, open source",
      "skills": [
        "local-ai",
        "ai-evaluation",
        "product-strategy"
      ],
      "lastActivity": "2026-11-06",
      "activityCheckedOn": "2026-11-05",
      "links": [
        {
          "label": "Docs",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/"
        },
        {
          "label": "GitHub",
          "href": "https://github.com/opnsrcntrbtr/lecture-lens"
        }
      ],
      "results": {
        "title": "Quality, limits and costs",
        "asOf": "Prototype full eval run, 14 Oct 2026 · project documentation, Oct 2026",
        "note": "Tick = the 3 Oct baseline, before the judge scored the answer body only; same generator and prompt versions. Eval run summaries are in the private prototype; the harness and statistics are public. Transcript completeness is from one measured live class.",
        "source": {
          "label": "Evaluation guide",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/guides/evaluation"
        },
        "metrics": [
          {
            "kind": "rate",
            "label": "Answers faithful (43 of 43)",
            "value": 1,
            "floor": 0.86
          },
          {
            "kind": "rate",
            "label": "Answers relevant (41 of 43)",
            "value": 0.953,
            "floor": 0.721
          },
          {
            "kind": "rate",
            "label": "Flashcard quality (10 of 15)",
            "value": 0.667,
            "weak": true
          },
          {
            "kind": "rate",
            "label": "Words captured (one class)",
            "value": 0.75,
            "weak": true,
            "display": "~75%"
          },
          {
            "kind": "stat",
            "label": "Bytes uploaded",
            "value": "0"
          },
          {
            "kind": "stat",
            "label": "Build time, 90-min class",
            "value": "~10 min"
          }
        ]
      }
    },
    {
      "name": "CodeSearchNet Find & Explain",
      "url": "https://ainpm.in/work/codesearchnet/",
      "summary": "Fine-tuned MiniLM retrieval (MRR 0.815 vs BM25 0.527) and a fine-tuned flan-t5 summariser, evaluated on the official CodeSearchNet test split with significance tests.",
      "status": "Live demo",
      "skills": [
        "ai-evaluation",
        "rag-product",
        "communicating-quality"
      ],
      "lastActivity": "2026-08-13",
      "activityCheckedOn": "2026-11-05",
      "links": [
        {
          "label": "Live demo",
          "href": "https://huggingface.co/spaces/opnsrcntrbtrian/csne-find-and-explain"
        },
        {
          "label": "GitHub",
          "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain"
        },
        {
          "label": "Model",
          "href": "https://huggingface.co/opnsrcntrbtrian/csne-minilm-retrieval-full-1ep"
        }
      ],
      "results": {
        "title": "Retrieval on 1,000 test queries (1 gold + 999 distractors)",
        "asOf": "Project report, Aug 2026 · official CodeSearchNet Python test split",
        "note": "Tick = baseline (BM25 or zero-shot). Queries are author-written docstrings, which likely overstates real-world search quality.",
        "source": {
          "label": "REPORT.md",
          "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
        },
        "metrics": [
          {
            "kind": "rate",
            "label": "MRR, fine-tuned (BM25 0.527)",
            "value": 0.8151,
            "floor": 0.5272
          },
          {
            "kind": "rate",
            "label": "Recall@1",
            "value": 0.725,
            "floor": 0.407
          },
          {
            "kind": "rate",
            "label": "Recall@10",
            "value": 0.961,
            "floor": 0.747
          },
          {
            "kind": "rate",
            "label": "Summary BERTScore F1 (zero-shot 0.747)",
            "value": 0.8463,
            "floor": 0.7473
          }
        ]
      }
    }
  ],
  "contributions": {
    "checkedOn": "2026-11-05",
    "items": [
      {
        "kind": "merged-pr",
        "title": "Fix #563: prevent rationale-node leakage and preserve calls direction",
        "project": "Graphify-Labs/graphify",
        "href": "https://github.com/Graphify-Labs/graphify/pull/576",
        "date": "2026-05-02",
        "summary": "Fixed two bugs that inflated node importance and flipped call directions in a code knowledge-graph tool used with AI coding assistants. Phantom edges in the repro fell from 7 to 0, with 5 new regression tests (441 to 446 passing).",
        "skills": [
          "ai-evaluation",
          "communicating-quality"
        ]
      },
      {
        "kind": "merged-pr",
        "title": "Add inline variant to the Filters component",
        "project": "o2sdev/openselfservice",
        "href": "https://github.com/o2sdev/openselfservice/pull/365",
        "date": "2025-12-10",
        "summary": "Added an inline layout option to a customer-portal UI component, backward compatible with the drawer layout and documented with Storybook stories.",
        "skills": [
          "program-delivery"
        ]
      },
      {
        "kind": "package",
        "title": "express-swagger-auto",
        "project": "npm",
        "href": "https://www.npmjs.com/package/express-swagger-auto",
        "date": "2025-11-22",
        "summary": "Generates OpenAPI 3.x specs and Swagger UI from Express APIs via decorators, JSDoc and runtime capture. v0.3.2.",
        "stat": "45 npm downloads, 5 Oct to 3 Nov 2026",
        "skills": [
          "program-delivery"
        ]
      },
      {
        "kind": "package",
        "title": "chroncraft",
        "project": "npm",
        "href": "https://www.npmjs.com/package/chroncraft",
        "date": "2025-11-27",
        "summary": "Lightweight, tree-shakeable date and time library with time-zone handling. v0.3.0.",
        "stat": "20 npm downloads, 5 Oct to 3 Nov 2026",
        "skills": [
          "program-delivery"
        ]
      },
      {
        "kind": "package",
        "title": "express-middleware-chain",
        "project": "npm",
        "href": "https://www.npmjs.com/package/express-middleware-chain",
        "date": "2025-12-07",
        "summary": "Fluent, typed middleware composition for Express with validation, rate limiting and caching. v0.1.0.",
        "stat": "16 npm downloads, 5 Oct to 3 Nov 2026",
        "skills": [
          "program-delivery"
        ]
      },
      {
        "kind": "package",
        "title": "password-suite",
        "project": "npm",
        "href": "https://www.npmjs.com/package/password-suite",
        "date": "2025-11-27",
        "summary": "Cryptographically secure password and passphrase generation with strength analysis, on the Web Crypto API. v2.0.1.",
        "stat": "14 npm downloads, 5 Oct to 3 Nov 2026",
        "skills": [
          "privacy-security"
        ]
      },
      {
        "kind": "package",
        "title": "express-type-safe-routes",
        "project": "npm",
        "href": "https://www.npmjs.com/package/express-type-safe-routes",
        "date": "2025-12-07",
        "summary": "Type-safe Express routes with Zod inference for params, query, body and responses. v0.1.0.",
        "stat": "8 npm downloads, 5 Oct to 3 Nov 2026",
        "skills": [
          "program-delivery"
        ]
      },
      {
        "kind": "package",
        "title": "express-env-validator",
        "project": "npm",
        "href": "https://www.npmjs.com/package/express-env-validator",
        "date": "2025-12-08",
        "summary": "Fail-fast, typed environment configuration checks for Express apps. v0.1.0.",
        "stat": "7 npm downloads, 5 Oct to 3 Nov 2026",
        "skills": [
          "privacy-security"
        ]
      },
      {
        "kind": "package",
        "title": "pandera-unified-validator",
        "project": "PyPI",
        "href": "https://pypi.org/project/pandera-unified-validator/",
        "date": "2025-11-28",
        "summary": "Unifies Pydantic record validation with Pandera DataFrame checks across pandas and Polars. v0.1.0.",
        "stat": "11 PyPI downloads in the month to 5 Nov 2026",
        "skills": [
          "ai-evaluation"
        ]
      }
    ]
  },
  "experiments": [
    {
      "name": "MLX AutoResearch agent",
      "href": "https://github.com/opnsrcntrbtr/mlx-autoresearch-agent",
      "started": "2026-08-04",
      "lastActivity": "2026-08-05",
      "what": "A local agent (Qwen on MLX) edits one training script, trains for 5 minutes on an Apple silicon Mac, and keeps the change only if validation loss improves.",
      "status": "Harness and loop spec published; no results yet",
      "skills": [
        "local-ai",
        "ai-evaluation"
      ]
    },
    {
      "name": "SEBI regulatory knowledge-base agent",
      "href": "https://github.com/opnsrcntrbtr/sebi-regulatory-knowledge-base-ai-agent",
      "started": "2025-11-27",
      "lastActivity": "2025-11-28",
      "what": "An agentic RAG toolkit over SEBI regulations in LangChain.js and LangGraph; an early step before SEBI Circular RAG.",
      "status": "Prototype, superseded",
      "skills": [
        "rag-product",
        "trust-compliance"
      ]
    },
    {
      "name": "UPI offline payment PWA (NFC)",
      "href": "https://github.com/opnsrcntrbtr/UPI-Offline-Payment-PWA-NFC-Based",
      "started": "2025-11-15",
      "lastActivity": "2025-11-28",
      "what": "Design for offline UPI payments by NFC tap with QR fallback, enforcing UPI Lite limits (Rs 1,000 per payment, Rs 5,000 balance).",
      "status": "Scaffolding only; payment flows not built",
      "skills": [
        "trust-compliance",
        "privacy-security"
      ]
    },
    {
      "name": "GenAI Tasker Lite (Android)",
      "href": "https://github.com/opnsrcntrbtr/genAI-edge-tasker-lite",
      "started": "2025-09-13",
      "lastActivity": "2026-05-31",
      "what": "An offline-first Android 14 task manager with on-device semantic search, plus a phased roadmap to real embeddings.",
      "status": "Educational scaffold",
      "skills": [
        "local-ai"
      ]
    }
  ],
  "productArtifacts": [
    {
      "slug": "sebi-rag",
      "name": "SEBI Circular RAG",
      "caseStudy": "/sebi-rag/",
      "updated": "2026-11-08",
      "brief": {
        "problem": "SEBI amends circulars and folds them into master circulars. An answer built on a superseded circular can look right and still be non-compliant, and generic chatbots invent circular numbers.",
        "users": "Compliance officers and product managers at mutual fund AMCs, brokers, depositories, RTAs and fintechs; RegTech builders; Indian legal-NLP researchers.",
        "job": "Find the circular that governs this, show the exact text, say whether it still applies, and make nothing up.",
        "constraints": [
          "Local-first on Apple silicon; official SEBI publications are the only authority",
          "When evidence is insufficient, answer \"I don’t know based on the available evidence\"",
          "One pinned benchmark runtime, model versions and seeds, so results are reproducible",
          "Optimise only validated stages, and only for a measurable benefit (default bar: 10%)"
        ],
        "nonGoals": [
          "Legal advice or unsupported legal conclusions",
          "API-only proprietary models that would send SEBI queries to a third party",
          "Targeting the Apple Neural Engine (declined in ADR-003)"
        ],
        "source": {
          "label": "project_context.md §1, §8, §9",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/project_context.md"
        }
      },
      "evalSpec": {
        "summary": "A 260-question adjudicated golden set, stratified by failure type, with six release floors. A change that drops any metric below its floor does not ship.",
        "dataset": [
          "golden_v7: 260 adjudicated questions; 41 should be declined; 15 ask for the rule as of a date",
          "Strata: 40 title lookups, 60 paraphrases, 30 numeric or table, 40 supersession lineage, 20 multi-hop, 20 repealed basis, 40 hard negatives, 10 far negatives",
          "Adjudicated with a local 35B model, with agreement checks and escalations"
        ],
        "metrics": [
          {
            "name": "Abstention accuracy",
            "value": "0.981",
            "floor": "0.937"
          },
          {
            "name": "Recall@10",
            "value": "0.943",
            "floor": "0.840"
          },
          {
            "name": "Context recall",
            "value": "0.916",
            "floor": "0.819"
          },
          {
            "name": "Citation recall",
            "value": "0.881",
            "floor": "0.735"
          },
          {
            "name": "nDCG@10",
            "value": "0.697",
            "floor": "0.593"
          },
          {
            "name": "Citation precision",
            "value": "0.194",
            "floor": "0.147",
            "note": "Known weak spot"
          }
        ],
        "rule": "Floors belong to the stack, not the corpus: changing the generator, embedder, corpus size, chunker version or citation margin invalidates them, and they must be re-derived before comparing anything.",
        "source": {
          "label": "project_context.md §7.4 and the evaluation log",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/status.md"
        }
      },
      "metricsTree": {
        "northStar": "Questions answered correctly with a checkable citation, or honestly declined",
        "why": "In regulation a confident wrong answer costs more than no answer, so \"honestly declined\" counts as success.",
        "drivers": [
          {
            "name": "Find the right circular",
            "metric": "Recall@10 0.943 · nDCG@10 0.697"
          },
          {
            "name": "Use the right passage",
            "metric": "Context recall 0.916"
          },
          {
            "name": "Cite what was used",
            "metric": "Citation recall 0.881"
          }
        ],
        "guardrails": [
          {
            "name": "Decline when evidence is missing",
            "metric": "Abstention accuracy 0.981"
          },
          {
            "name": "Don’t over-cite",
            "metric": "Citation precision 0.194 (weak)"
          },
          {
            "name": "Never cite a forbidden source",
            "metric": "must_not_cite violation rate"
          }
        ]
      },
      "decisions": [
        {
          "id": "D1",
          "date": "2026-07-02",
          "title": "Hybrid retrieval is mandatory",
          "decision": "Dense (FAISS) plus sparse (BM25) retrieval fused with RRF; a new dense store only on 10% or better evidence.",
          "why": "The retrieval stage is fixed architecture; the model inside it is replaceable only on measured evidence.",
          "status": "Amended",
          "source": {
            "label": "ADR-001",
            "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/adr-001-architecture-review-2026-07.md"
          }
        },
        {
          "id": "ADR-002",
          "date": "2026-07-02",
          "title": "Certainty bands, not probabilities",
          "decision": "Every answer carries high, medium or low certainty and a reason when it declines; a labelled low-confidence draft only on explicit opt-in.",
          "why": "Raw model scores are not certainty (a hard negative scored 0.985), and 56 golden items were too few to fit a calibrator.",
          "status": "Accepted",
          "source": {
            "label": "ADR-002",
            "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/adr-002-certainty-and-advisory.md"
          }
        },
        {
          "id": "ADR-003",
          "date": "2026-07-22",
          "title": "Apple Neural Engine declined",
          "decision": "Keep generation on MLX and embeddings and reranking on MPS/MLX.",
          "why": "The Neural Engine trades speed for energy efficiency and needs a Core ML conversion; a plugged-in server pipeline needs throughput, where MLX on the GPU wins.",
          "status": "Declined",
          "source": {
            "label": "ADR-003",
            "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/adr-003-ane-declined.md"
          }
        },
        {
          "id": "D2",
          "date": "2026-08-20",
          "title": "Proprietary API embedder rejected",
          "decision": "Voyage models rejected as candidates, regardless of benchmark strength.",
          "why": "API-only service conflicts with local-first and would send SEBI queries to a third party.",
          "status": "Rejected",
          "source": {
            "label": "project_context.md §8",
            "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/project_context.md"
          }
        },
        {
          "id": "ADR-004",
          "date": "2026-08-24",
          "title": "New reranker adopted below the 10% bar",
          "decision": "jina-reranker-v3 adopted by explicit owner override: recall@10 +2.42% and nDCG@10 +6.76% on SEBI data, no regression.",
          "why": "A consistent, non-regressive gain; the 10% bar was self-imposed, and the override and its reasoning are recorded.",
          "status": "Adopted",
          "source": {
            "label": "ADR-004",
            "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/adr-004-reranker-candidate-reassessment-2026-08.md"
          }
        }
      ],
      "launch": {
        "status": "Live demo on Hugging Face; open code and dataset",
        "shipped": [
          "Index of 1,490 circulars and 83,752 chunks (15 Sep 2026)",
          "Answers with numbered citations, supersession status and successor circulars",
          "Certainty block and decline reason on every response",
          "Open dataset of SEBI circulars on Hugging Face (CC BY 4.0)"
        ],
        "limits": [
          "Citation precision 0.194: answers cite more passages than they need",
          "The armed gate file has no stack fingerprint, so drift is flagged as \"unverifiable\" until floors are re-derived",
          "English circulars only"
        ],
        "next": [
          "Add a stack fingerprint to the gate and re-derive floors on the current chunker",
          "Improve citation precision without losing citation recall"
        ],
        "source": {
          "label": "Evaluation log (status.md)",
          "href": "https://github.com/opnsrcntrbtr/sebi-circular-rag/blob/main/docs/status.md"
        }
      },
      "url": "https://ainpm.in/work/artifacts/sebi-rag/"
    },
    {
      "slug": "codesearchnet",
      "name": "CodeSearchNet Find & Explain",
      "caseStudy": "/work/codesearchnet/",
      "updated": "2026-11-08",
      "brief": {
        "problem": "Keyword search fails when a developer describes intent and the code uses different words; docstrings and code share little vocabulary.",
        "users": "Developers searching unfamiliar Python codebases.",
        "job": "Describe what a function does, see the best matches, and understand each one without reading its body.",
        "constraints": [
          "Small enough to fine-tune on a free Colab T4 GPU",
          "Official CodeSearchNet splits, already separated by repository, so test code never leaks into training",
          "Exact search, not approximate, so index error cannot confound the metrics",
          "Both summariser backends get the identical prompt, so they are comparable"
        ],
        "nonGoals": [
          "Languages other than Python",
          "Approximate-nearest-neighbour indexing at this corpus size",
          "Real developer query logs (none were available)"
        ],
        "source": {
          "label": "REPORT.md §1, §3, §4",
          "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
        }
      },
      "evalSpec": {
        "summary": "Each of 1,000 test queries is ranked against its gold function plus 999 distractors; lexical and zero-shot baselines set the bar, and gains must survive significance tests.",
        "dataset": [
          "Official CodeSearchNet Python test split; 393,427 training pairs after filtering and de-duplication",
          "A gold function never ranked counts as a miss (0), not dropped",
          "Summaries scored on 50 test examples with BLEU and BERTScore"
        ],
        "metrics": [
          {
            "name": "MRR",
            "value": "0.815",
            "floor": "0.527 (BM25)"
          },
          {
            "name": "Recall@1",
            "value": "0.725",
            "floor": "0.407 (BM25)"
          },
          {
            "name": "Recall@10",
            "value": "0.961",
            "floor": "0.747 (BM25)"
          },
          {
            "name": "nDCG@10",
            "value": "0.850",
            "floor": "0.574 (BM25)"
          },
          {
            "name": "Summary BERTScore F1",
            "value": "0.846",
            "floor": "0.747 (zero-shot)"
          }
        ],
        "rule": "A claimed gain needs a paired bootstrap (10,000 resamples) and a paired t-test; the 50K-to-393K data gain of +0.022 MRR passed both (p < 0.001), so it is reported as real but small.",
        "source": {
          "label": "REPORT.md §4.3, §5",
          "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
        }
      },
      "metricsTree": {
        "northStar": "The right function appears in the first results and is understood without reading it",
        "why": "Search only helps if the answer is near the top and the developer can judge it quickly.",
        "drivers": [
          {
            "name": "Right function near the top",
            "metric": "Recall@10 0.961 · MRR 0.815"
          },
          {
            "name": "Explanation matches meaning",
            "metric": "BERTScore F1 0.846"
          }
        ],
        "guardrails": [
          {
            "name": "Holds up in a bigger search space",
            "metric": "MRR 0.755 at 4,999 distractors (0.882 at 99)"
          },
          {
            "name": "No empty explanations",
            "metric": "1 of 50 empty, replaced with a documented sentinel"
          },
          {
            "name": "Realistic queries",
            "metric": "Docstring queries flagged as likely optimistic"
          }
        ]
      },
      "decisions": [
        {
          "id": "Data",
          "date": "2026-08-10",
          "title": "Join two datasets on the GitHub permalink",
          "decision": "Use CodeSearchNet plus a dataset with curated one-line summaries; 100% join rate on a 20,000-row sample.",
          "why": "The simpler alternative lacked repository and language metadata, so Python filtering and repo-disjoint splits were impossible.",
          "status": "Adopted",
          "source": {
            "label": "REPORT.md §3.1",
            "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
          }
        },
        {
          "id": "Prep",
          "date": "2026-08-10",
          "title": "Strip docstrings from the code side",
          "decision": "Remove docstrings from code and truncate query docstrings at the first structured section.",
          "why": "Leaving them in makes retrieval trivially lexical.",
          "status": "Adopted",
          "source": {
            "label": "REPORT.md §3.3",
            "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
          }
        },
        {
          "id": "Index",
          "date": "2026-08-10",
          "title": "Exact search instead of approximate",
          "decision": "Brute-force search over normalised embeddings.",
          "why": "Approximation error would confound the metrics at this corpus size.",
          "status": "Adopted",
          "source": {
            "label": "REPORT.md §4.1",
            "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
          }
        },
        {
          "id": "Model",
          "date": "2026-08-10",
          "title": "Smaller summariser rejected",
          "decision": "flan-t5-small (80M) dropped for flan-t5-base (250M).",
          "why": "It could not follow the prompt and degenerated into repeated unknown tokens.",
          "status": "Rejected",
          "source": {
            "label": "REPORT.md §4.2",
            "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
          }
        },
        {
          "id": "Next",
          "date": "2026-08-10",
          "title": "More epochs before more data",
          "decision": "The next experiment is a 3-epoch run, not a larger corpus.",
          "why": "7.9 times more data added only 0.022 MRR.",
          "status": "Accepted",
          "source": {
            "label": "REPORT.md §5.3",
            "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
          }
        }
      ],
      "launch": {
        "status": "Live demo on Hugging Face Spaces; models and code public",
        "shipped": [
          "Gradio demo with retrieval-only and find-and-explain modes",
          "Two public retrieval models (50K and full corpus)",
          "Full report with ablation, pool-size sensitivity and error analysis"
        ],
        "limits": [
          "Docstring-as-query bias likely overstates real-world quality",
          "Python only; one epoch for every model",
          "Summaries scored on 50 examples"
        ],
        "next": [
          "3-epoch fine-tune on the full corpus",
          "Evaluation on real developer queries",
          "Other languages and an approximate index for larger corpora"
        ],
        "source": {
          "label": "REPORT.md §7, §8",
          "href": "https://github.com/opnsrcntrbtr/codesearchnet-find-and-explain/blob/main/REPORT.md"
        }
      },
      "url": "https://ainpm.in/work/artifacts/codesearchnet/"
    },
    {
      "slug": "lecture-lens",
      "name": "Lecture Lens",
      "caseStudy": "/work/lecture-lens/",
      "updated": "2026-11-11",
      "brief": {
        "problem": "Working professionals in part-time online courses attend recorded lectures and live classes with little time to re-watch either, and cloud note-takers upload course material.",
        "users": "A working professional in a part-time online course, on an Apple silicon Mac.",
        "job": "When I attend a class, I want it captured and turned into material I can trust and will remember, without watching the tooling.",
        "constraints": [
          "Everything runs on the Mac; nothing is uploaded",
          "Copy-protected video is never recorded; audio is kept",
          "One user, no Apple Developer certificate",
          "Evaluation must never disrupt a live class"
        ],
        "nonGoals": [
          "Recording everything and filtering later",
          "Longer notes and weekly digests",
          "Hosted evaluation dashboards"
        ],
        "source": {
          "label": "Roadmap and ADR-002",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/project/roadmap"
        }
      },
      "evalSpec": {
        "summary": "Every prompt or model change can be measured locally, with a judge model that differs from the generator and is itself benchmarked.",
        "dataset": [
          "Smoke: 6 fixed cases (2 questions, 2 notes, 2 card sets), about 12 minutes",
          "Full: all cases, scored with DeepEval faithfulness, answer relevancy and G-Eval criteria, over an hour",
          "Judge benchmark on labelled claim and context pairs; synthetic slides with exact answers"
        ],
        "metrics": [
          {
            "name": "Words captured, one measured class",
            "value": "~75%",
            "note": "Weak spot"
          },
          {
            "name": "Bytes uploaded",
            "value": "0"
          },
          {
            "name": "Build time for a 90-minute class",
            "value": "~10 min"
          },
          {
            "name": "Smoke-eval cases",
            "value": "6"
          },
          {
            "name": "Notes cases passed (prototype, 3 Oct 2026)",
            "value": "15 of 15",
            "floor": "Faithfulness ≥ 0.85"
          },
          {
            "name": "Answer cases passing every metric (3 Oct baseline)",
            "value": "26 of 43",
            "floor": "Faithfulness ≥ 0.8"
          },
          {
            "name": "Answers faithful / relevant after the judge fix (14 Oct)",
            "value": "43 and 41 of 43",
            "floor": "Faithfulness ≥ 0.8, relevancy ≥ 0.7"
          },
          {
            "name": "Flashcard cases passing every metric (3 Oct)",
            "value": "9 of 15",
            "floor": "Faithfulness ≥ 0.85",
            "note": "Weak spot"
          },
          {
            "name": "Flashcard quality (14 Oct)",
            "value": "10 of 15",
            "floor": "G-Eval ≥ 0.7",
            "note": "Weak spot"
          },
          {
            "name": "Judge agreement with labels, blind holdout",
            "value": "80%, kappa 0.44"
          }
        ],
        "rule": "Check what a failure is before fixing the model: the first full baseline failed mostly because the judge penalised a source list it could not see, so only the answer body is scored now.",
        "source": {
          "label": "Evaluation guide",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/guides/evaluation"
        }
      },
      "metricsTree": {
        "northStar": "Classes captured and turned into notes and cards the learner trusts and reviews",
        "why": "Capture that misses the class, or notes the learner cannot trust, make the rest useless.",
        "drivers": [
          {
            "name": "Capture completeness",
            "metric": "Words captured ~75% in one measured class"
          },
          {
            "name": "Notes stay faithful",
            "metric": "DeepEval faithfulness and answer relevancy"
          },
          {
            "name": "Cards get reviewed",
            "metric": "Flashcard review log (planned)"
          }
        ],
        "guardrails": [
          {
            "name": "Privacy",
            "metric": "0 bytes uploaded"
          },
          {
            "name": "Respect protected video",
            "metric": "Screen capture pauses on DRM players"
          },
          {
            "name": "Don’t break a live class",
            "metric": "Evaluations refuse to run during capture"
          }
        ]
      },
      "decisions": [
        {
          "id": "Vision",
          "date": "2026-09-29",
          "title": "Slide descriptions chosen on realistic frames",
          "decision": "Use a 35B multimodal model with thinking off, crop to the lecture player, and discard any description where the image provably never reached the model.",
          "why": "Thinking on looked safer on clean slides but passed only 15 of 21 realistic ones; production-style input passed 21 of 21 (prototype ADR-001).",
          "status": "Accepted",
          "source": {
            "label": "Origin story: prototype evidence",
            "href": "https://ainpm.in/work/lecture-lens/origin/#proved"
          }
        },
        {
          "id": "ADR-002",
          "date": "2026-09-29",
          "title": "Native SwiftUI app launches the toolkit",
          "decision": "A menu-bar app starts capture itself and calls the tools through a JSON contract.",
          "why": "macOS grants screen recording to the app that launches capture; terminal launches produced 0 frames.",
          "status": "Accepted",
          "source": {
            "label": "ADR-002",
            "href": "https://opnsrcntrbtr.github.io/lecture-lens/decisions/ADR-002-native-app-integration"
          }
        },
        {
          "id": "ADR-004",
          "date": "2026-09-30",
          "title": "Pause screen capture on protected video, keep audio",
          "decision": "Pause capture on listed DRM player pages and build sessions from page markers and the transcript.",
          "why": "Protected video went black while capture ran; the protection is respected and the transcript is what matters.",
          "status": "Accepted",
          "source": {
            "label": "ADR-004",
            "href": "https://opnsrcntrbtr.github.io/lecture-lens/decisions/ADR-004-drm-video-black-screen"
          }
        },
        {
          "id": "ADR-005",
          "date": "2026-11-04",
          "title": "Confirm repeated slides with structural similarity",
          "decision": "Keep the cheap hash test, then confirm a repeat only if at most 5 of 36 grid cells changed.",
          "why": "Slides from one template look alike to the hash, so the old test missed most of them (recall 0.17 to 0.58 on synthetic decks); counting changed cells reached recall 1.00. Accepted on synthetic evidence, to confirm on a real build.",
          "status": "Accepted",
          "source": {
            "label": "ADR-005",
            "href": "https://opnsrcntrbtr.github.io/lecture-lens/decisions/ADR-005-keyframe-ssim"
          }
        },
        {
          "id": "ADR-006",
          "date": "2026-11-05",
          "title": "Run the language model inside Chrome",
          "decision": "Add a browser provider: Gemini Nano first, a WebGPU model as fallback; the local model server stays the default.",
          "why": "The local server costs 20 to 40 GB of weights and a capable Mac; the browser needs neither.",
          "status": "Accepted",
          "source": {
            "label": "ADR-006",
            "href": "https://opnsrcntrbtr.github.io/lecture-lens/decisions/ADR-006-in-browser-llm"
          }
        }
      ],
      "launch": {
        "status": "Early, open source; Lecture Lens Studio in MVP",
        "shipped": [
          "SwiftUI menu-bar app over a zsh and Python toolkit",
          "Local evaluation harness with a calibrated judge",
          "Six-layer content guard for a public repo",
          "Studio: notes, cards, answers, live Q&A and in-browser transcription inside Chrome (5 Nov 2026)"
        ],
        "limits": [
          "macOS on Apple silicon only",
          "About a quarter of spoken words missing in one measured class",
          "A full evaluation takes over an hour",
          "Depends on screenpipe, which it does not control"
        ],
        "next": [
          "Capture guard that alerts within 60 seconds when nothing is saved",
          "Cancellable evaluations that show the change against the last run",
          "Eval checks drawn from real failures"
        ],
        "source": {
          "label": "Roadmap",
          "href": "https://opnsrcntrbtr.github.io/lecture-lens/project/roadmap"
        }
      },
      "url": "https://ainpm.in/work/artifacts/lecture-lens/"
    },
    {
      "slug": "trustvault",
      "name": "TrustVault PWA",
      "caseStudy": "/work/trustvault/",
      "updated": "2026-11-08",
      "brief": {
        "problem": "Password managers ask users to trust a server with their vault. A progressive web app can keep every secret on the device, but only if storage alone can never unlock it.",
        "users": "Privacy-conscious individuals and families who want a password manager that works offline on any device.",
        "job": "Keep my passwords safe on my own device, unlock with my fingerprint, and never depend on a company’s server.",
        "constraints": [
          "Zero-knowledge: a full storage dump must not reveal the vault key",
          "Offline-first; works as an installable PWA",
          "OWASP Mobile Top 10 (2025) as the security baseline",
          "Any AI help runs locally, and no secret is ever sent to a model"
        ],
        "nonGoals": [
          "Cloud sync of vault contents",
          "Telemetry or hidden features",
          "Remote AI services"
        ],
        "source": {
          "label": "PROJECT_STATUS.md (6 Jul 2026)",
          "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
        }
      },
      "evalSpec": {
        "summary": "Security and quality are gated by an automated suite plus OWASP and Lighthouse checks. Figures are self-reported by the project; there is no independent audit yet.",
        "dataset": [
          "About 600 unit, 400 integration and 100 security tests (crypto, auth flow, CRUD, OWASP checks)",
          "Lighthouse audits for performance and PWA installability",
          "OWASP Mobile Top 10 (2025) checklist, M1 to M10"
        ],
        "metrics": [
          {
            "name": "Automated tests passing",
            "value": "4,139 of 4,142",
            "floor": ">85% coverage target"
          },
          {
            "name": "TypeScript errors",
            "value": "0",
            "floor": "0"
          },
          {
            "name": "Lighthouse PWA",
            "value": "100",
            "floor": ">90"
          },
          {
            "name": "ESLint errors in src/",
            "value": "801",
            "floor": "0",
            "note": "Open debt"
          },
          {
            "name": "WCAG 2.1 AA",
            "value": "Partial",
            "floor": "Formal audit",
            "note": "Pending"
          }
        ],
        "rule": "Status is reported against targets, including the misses: lint errors and the accessibility audit are listed as open, not hidden.",
        "source": {
          "label": "PROJECT_STATUS.md: codebase health and tests",
          "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
        }
      },
      "metricsTree": {
        "northStar": "Vaults that stay usable and unreadable to anyone but their owner",
        "why": "A password manager fails if it leaks, and also if its owner gets locked out.",
        "drivers": [
          {
            "name": "Fast, safe unlock",
            "metric": "Passkey (WebAuthn PRF) unlock with a password fallback"
          },
          {
            "name": "Works anywhere",
            "metric": "Lighthouse PWA 100; offline-first caching"
          },
          {
            "name": "Quality holds",
            "metric": "4,139 of 4,142 tests passing"
          }
        ],
        "guardrails": [
          {
            "name": "Storage alone never unlocks",
            "metric": "Wrap key never stored; non-extractable session keys"
          },
          {
            "name": "No data egress",
            "metric": "Zero telemetry; local-only AI with secret-free prompts"
          },
          {
            "name": "Breach checks leak nothing",
            "metric": "HIBP k-anonymity (5-character hash prefix)"
          }
        ]
      },
      "decisions": [
        {
          "id": "Phase 4",
          "date": "2026-06-10",
          "title": "Passkey unlock without a stored key",
          "decision": "Unlock with a key derived from the authenticator’s PRF output, released only after biometric or PIN verification and never stored.",
          "why": "Zero-knowledge requires that nothing persisted can reconstruct the vault key.",
          "status": "Adopted",
          "source": {
            "label": "PROJECT_STATUS.md Phase 4",
            "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
          }
        },
        {
          "id": "Phase 6",
          "date": "2026-06-15",
          "title": "Strict CSP and no third-party CDNs",
          "decision": "Hash-based Content Security Policy with no unsafe-inline, and self-hosted OCR instead of a CDN.",
          "why": "Every external script is a way in; zero CDN egress keeps the trust boundary on the device.",
          "status": "Adopted",
          "source": {
            "label": "PROJECT_STATUS.md Phase 6 and security posture",
            "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
          }
        },
        {
          "id": "Phase 6",
          "date": "2026-06-15",
          "title": "Breach checks by k-anonymity",
          "decision": "Query Have I Been Pwned with only a 5-character SHA-1 prefix.",
          "why": "Users get breach alerts without the password or its full hash leaving the device.",
          "status": "Adopted",
          "source": {
            "label": "PROJECT_STATUS.md Phase 6",
            "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
          }
        },
        {
          "id": "AI",
          "date": "2026-06-21",
          "title": "On-device AI behind a zero-egress boundary",
          "decision": "Gemini Nano on desktop and WebLLM on Android explain password strength and breach impact locally; prompts never include secrets.",
          "why": "AI help is useful only if it cannot become a leak.",
          "status": "Adopted",
          "source": {
            "label": "PROJECT_STATUS.md AI assistance",
            "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
          }
        },
        {
          "id": "Library",
          "date": "2026-07-04",
          "title": "Extract the crypto into a reusable library",
          "decision": "Move PRF key wrapping into webauthn-prf-zktv, published on npm with a research preprint.",
          "why": "A standalone, tested library can be reviewed and reused independently of the app.",
          "status": "Adopted",
          "source": {
            "label": "webauthn-prf-zktv README",
            "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/README.md"
          }
        }
      ],
      "launch": {
        "status": "Staging deployment; feature-complete through Phase 7 plus experimental AI",
        "shipped": [
          "Credential management, generator, TOTP, auto-lock and encrypted import/export",
          "Passkey unlock and multi-vault profiles",
          "Breach detection and local AI explanations"
        ],
        "limits": [
          "801 ESLint errors in src/ to clear",
          "One flaky import-export test",
          "WCAG 2.1 AA formal audit pending",
          "No independent security audit yet"
        ],
        "next": [
          "Full lint sweep before the next feature",
          "Formal accessibility audit",
          "CSV import/export and an error-boundary UI"
        ],
        "source": {
          "label": "PROJECT_STATUS.md: known limitations",
          "href": "https://github.com/opnsrcntrbtr/TrustVault-PWA/blob/main/docs/status/PROJECT_STATUS.md"
        }
      },
      "url": "https://ainpm.in/work/artifacts/trustvault/"
    },
    {
      "slug": "webauthn-prf-zktv",
      "name": "webauthn-prf-zktv",
      "caseStudy": "/work/webauthn-prf-zktv/",
      "updated": "2026-11-08",
      "brief": {
        "problem": "Web apps that encrypt user data need a key that a stolen database cannot reproduce. Passkeys can supply one through the WebAuthn PRF extension, but the pattern is easy to get wrong.",
        "users": "Developers building zero-knowledge web and PWA apps; researchers reviewing passkey-based encryption.",
        "job": "Give me a small, reviewed library that wraps a vault key under a passkey or a password, so storage alone can never unlock it.",
        "constraints": [
          "One runtime dependency; native WebAuthn APIs, no wrapper library",
          "Only non-extractable keys; nothing secret is ever persisted",
          "Symmetric primitives only in the wrap path (AES-256-GCM, HKDF-SHA256, scrypt)"
        ],
        "nonGoals": [
          "Combined-factor (hybrid) wrapping in v1",
          "Android WebView PRF, which the platform does not support"
        ],
        "source": {
          "label": "README and SECURITY.md",
          "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/SECURITY.md"
        }
      },
      "evalSpec": {
        "summary": "Correctness is pinned by a threat model with explicit guarantees, known-answer interop vectors for independent implementations, and an automated test suite.",
        "dataset": [
          "Primary threat: an attacker with a full storage dump; secondary: script running in the page (XSS)",
          "Known-answer vectors so other implementations can check compatibility",
          "Vitest suite run by a verify script with type-check, lint and build"
        ],
        "metrics": [
          {
            "name": "Platforms with WebAuthn PRF",
            "value": "5 of 6",
            "note": "Android WebView unsupported"
          },
          {
            "name": "Runtime dependencies",
            "value": "1"
          },
          {
            "name": "Wrap schemes",
            "value": "2",
            "note": "Passkey and password"
          },
          {
            "name": "scrypt cost",
            "value": "N=131072, r=8, p=1"
          }
        ],
        "rule": "Storage compromise alone must never be sufficient: no stored input can recompute a key, and decrypt failures give no oracle.",
        "source": {
          "label": "SECURITY.md guarantees",
          "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/SECURITY.md"
        }
      },
      "metricsTree": {
        "northStar": "Apps can adopt passkey-based encryption without inventing their own crypto",
        "why": "The library only matters if others can use it safely.",
        "drivers": [
          {
            "name": "Works on the platforms people use",
            "metric": "PRF on 5 of 6 platforms"
          },
          {
            "name": "Easy to adopt",
            "metric": "npm v0.2.0, one dependency"
          },
          {
            "name": "Independently checkable",
            "metric": "Interop vectors and a DOI preprint"
          }
        ],
        "guardrails": [
          {
            "name": "No lockout",
            "metric": "Password wrap kept alongside every passkey wrap"
          },
          {
            "name": "No key export",
            "metric": "No exportKey path in the public API"
          },
          {
            "name": "No decrypt oracle",
            "metric": "One generic DecryptError"
          }
        ]
      },
      "decisions": [
        {
          "id": "Schemes",
          "date": "2026-07-05",
          "title": "Always keep a password wrap",
          "decision": "Treat the password scheme as the safety net that makes passkey unlock safe to deploy.",
          "why": "Deleting a passkey destroys its PRF path permanently, and the data with it.",
          "status": "Adopted",
          "source": {
            "label": "README: why two schemes",
            "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/README.md"
          }
        },
        {
          "id": "Deps",
          "date": "2026-07-05",
          "title": "Native WebAuthn, one dependency",
          "decision": "Call navigator.credentials directly with strict PRF validation; depend only on a scrypt implementation.",
          "why": "Fewer dependencies mean a smaller audit surface for a crypto library.",
          "status": "Adopted",
          "source": {
            "label": "README",
            "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/README.md"
          }
        },
        {
          "id": "Keys",
          "date": "2026-07-05",
          "title": "Non-extractable keys and a generic decrypt error",
          "decision": "No exportKey path; one DecryptError for every failure; zeroise transient buffers.",
          "why": "Script in the page can misuse an unlocked key but cannot steal it, and errors reveal nothing.",
          "status": "Adopted",
          "source": {
            "label": "SECURITY.md",
            "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/SECURITY.md"
          }
        },
        {
          "id": "Hybrid",
          "date": "2026-07-05",
          "title": "Combined-factor scheme deferred",
          "decision": "The hybrid-v1 name is reserved but not implemented in v1.",
          "why": "Ship the two well-understood schemes first.",
          "status": "Declined",
          "source": {
            "label": "SECURITY.md residuals",
            "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/SECURITY.md"
          }
        }
      ],
      "launch": {
        "status": "npm v0.2.0 (5 Jul 2026); preprint with DOI",
        "shipped": [
          "Core, WebAuthn and IndexedDB modules",
          "Threat model with guarantees and residuals",
          "Interop vectors and a CC BY 4.0 preprint"
        ],
        "limits": [
          "Zeroisation is best-effort in JavaScript",
          "PRF output passes through JS memory briefly, by spec",
          "No PRF in Android WebView; password scheme only there"
        ],
        "next": [
          "A combined-factor scheme",
          "Independent review of the implementation"
        ],
        "source": {
          "label": "SECURITY.md residuals",
          "href": "https://github.com/opnsrcntrbtr/webauthn-prf-zktv/blob/main/SECURITY.md"
        }
      },
      "url": "https://ainpm.in/work/artifacts/webauthn-prf-zktv/"
    }
  ]
}