SkillSync Co-Pilot

We don't replace the assessor. We help every assessor score like the best one, and make every certificate audit-proof, even offline in a village workshop.

AI-assisted, assessor-calibrated, offline-first Recognition of Prior Learning (RPL) assessment. Smart India Hackathon 2026 · Problem Statement 26242 · Ministry of Skill Development & Entrepreneurship. Demo trade: Welding (official NQR QP QG-4.5-AU-01811-2024-V1-ASDC), plus 10 drafted QPs for mapping.

Worker: voice declaration (Hindi)	Assessor: blind-first scoring

More screenshots, desktop and phone: docs/screenshots/.

What the problem statement asks → where it lives

PS 26242 asks for	Module	Where in the code
Structured self-declaration, auto-mapped to the closest NSQF QP(s)	M1 voice / typed / icon-card "Kaam ki Kahani" (Hindi, Hinglish, English) · M2 hybrid QP mapping with top-3, confidence and per-PC claim prefill	DeclareView.vue, speech.ts, mapping.py, lexicon (399 entries), on-device mapper-lite
Practical evaluation with guided checklists and image aids	M3 PC-level evidence plan (critical PCs locked to live observation) · M4 signed Process–Product–Person capture with a challenge code · M5 blind-first AI shadow rater	evidence_plan.py, CaptureView.vue, imageQc.ts, deviceKey.ts, integrity.py, shadow_rater.py, scoring.py
Standardised rubrics across assessors and locations	Rubric anchors on every PC · M6 hidden golden packets, Krippendorff's α, a private Calibration Index, drift heat-maps	rubric.py, RubricAnchors.vue, calibration.py, krippendorff.py, QaHome.vue
NSQF competency profile + recommendation for assessor sign-off	Per-PC profile, Full / Partial / Gap recommendation, bridge modules; sign-off with password re-entry → W3C Verifiable Credential + QR verify page	profile.py, credentials.py, credential.py, VerifyView.vue
Works in low connectivity	Installable PWA, service worker, encrypted on-device vault, outbox with idempotent sync, resumable chunked uploads	sw.ts, sync.ts, db.ts, vault.ts, evidence.ts
Evidence of consistency improvement	Simulated AB/BA crossover study, plus a live study mode in the app (assistance OFF, then ON)	simulate_study.py, studies.py, results
Where the tool supports vs replaces judgement	The table below, enforced in the API rather than the UI	scoring.py (blind-first), credentials.py (human-only sign-off)

Every important write lands in a hash-chained audit ledger (audit.py, GET /audit/verify) and a DPDP-style consent ledger (consent.py).

Architecture

flowchart LR
  subgraph Device["Field device · PWA, offline-first"]
    W["Worker<br/>voice · typed · icon cards"] --> ML["mapper-lite (TS)<br/>BM25 + lexicon"]
    A["Assessor<br/>checklist · P-P-P capture · blind scoring"] --> SG["QC + SHA-256<br/>+ ECDSA P-256 sign"]
    ML --> V[("AES-GCM vault + outbox<br/>IndexedDB")]
    SG --> V
  end
  subgraph API["FastAPI · JWT / RBAC"]
    GW["REST + signature verify"]
    M2["M2 QP mapping<br/>BM25 + lexicon + embeddings"]
    M3["M3 evidence plan"]
    INT["integrity<br/>pHash · EXIF · time · GPS"]
    M5["M5 shadow rater<br/>blind-first gate"]
    M6["M6 calibration α<br/>+ competency profile"]
  end
  subgraph Data["Data & audit"]
    DB[("SQLite · MySQL · SQL Server<br/>Postgres + pgvector")]
    FS[("local files · S3 store")]
    AL[("hash-chained ledger")]
  end
  subgraph Gate["Human gate"]
    SO["assessor sign-off"] --> VC["W3C VC + QR verify"]
  end
  V -- "outbox sync (idempotent)" --> GW
  GW -- "packs: QPs · lexicon" --> V
  GW --> M2 & M3 & INT & M5 & M6
  M2 & M3 & INT & M5 & M6 --> DB
  INT --> FS
  GW --> AL
  M6 --> SO

Loading

The full five-layer diagram and the key flows are in docs/ARCHITECTURE.md.

Quick start

Lite mode needs Python 3.11 and Node 20.19+ (or 22.12+). No Docker and no database server: it uses SQLite and local files.

git clone https://github.com/Aribhussain/skillsync-copilot && cd skillsync-copilot
npm run setup     # backend venv + deps, frontend deps, .env from .env.example
npm run lite      # API http://localhost:8000/docs  ·  PWA http://localhost:5173

On Windows, clone to a short path (for example C:\dev\skillsync): a few Python packages nest deeply and can exceed the 260-character path limit.

Optional: run npm run seed for a populated demo (6 SYNTHETIC workers, assessments in every state, golden packets, a signed credential). worker1 is left untouched for the live demo. On macOS / Linux, make setup, make lite-up and make seed do the same.

The optional multilingual embedding re-rank (about 1 GB, with torch) installs with backend/.venv/Scripts/pip install -r backend/requirements-embeddings.txt. Without it, BM25 + lexicon does the mapping.

Full mode (Docker): Postgres 16 + pgvector, SeaweedFS (S3), Valkey, a Celery worker, and nginx serving the PWA.

docker compose up -d --build --wait        # app http://localhost:8080 · API http://localhost:8000
backend/.venv/Scripts/python scripts/smoke_full.py   # end-to-end checks against the stack
docker compose down -v

MySQL 8 / SQL Server: python scripts/setup_db.py mysql --admin-user root creates the database and app user, runs the migrations and prints the DATABASE_URL for .env. The password is asked for at runtime and never stored. This path is compile-tested only; see Limitations.

Demo accounts

These are local demo defaults, shown on the login screen.

User	Role	Password
worker1 (Ramesh Kumar), worker2 (Salma Khatoon)	Worker	worker123
assessor1, assessor2, assessor3	Assessor	assessor123
qa1 (Dr. Kavita Rao)	QA lead	qa123
admin	Admin	admin123
worker3…worker8 (after npm run seed)	SYNTHETIC workers	worker123

5-minute demo for judges

The full script, with what to say at each step, is in docs/DEMO_SCRIPT.md.

1. worker1 → consent → speak or type the Hindi/Hinglish story → Welding is #1 with confidence, matched terms and pre-ticked PCs → confirm.
2. assessor1 (or assessor2/3: assignment picks the least-loaded) → the PC checklist, with critical PCs locked to live observation.
3. Turn the network off → guided Process → Product → Person capture with the one-time challenge code → the shots are signed and queued.
4. Score PCs blind; the AI opinion appears only after each lock. A big gap forces a reason code.
5. Network on → the outbox syncs and the server verifies every signature.
6. Upload the same photo for worker2 → it is flagged as reused.
7. Competency profile → Partial (GTAW gap → bridge module) → sign off with a password → a credential is issued → scan the QR → the verify page shows a valid signature.
8. qa1 → calibration dashboard: α, the assessor × NOS drift heat-map, and the audit chain verified.

Where the tool supports vs replaces judgement

Decision	AI's role	Final say
QP suggestion	Ranks the top 3 with confidence and matched terms	Worker + assessor confirm
Evidence plan	Rules engine; critical PCs locked to live observation	System rule
Evidence authenticity	Flags reuse, EXIF, time and GPS anomalies (flags only)	Assessor reviews
PC scores	Second opinion, revealed only after the blind score (server-enforced)	Assessor
Full / Partial / Gap	Recommends only	Assessor, with password re-entry and a justification when overriding
Certificate	None	Awarding body (the credential carries the assessor's decision)

Results

What	Result	Label
QP mapping, 30 held-out hand-written narratives (short, misspelled, cross-trade traps)	Top-3 93.3% (BM25 + lexicon), 96.7% (+ embeddings); top-1 90.0%	Hard set written before tuning. The narratives are not real worker data. Details
QP mapping, 60 template narratives	100% top-3	Sanity check only: the templates and lexicon share authors
Consistency, AB/BA crossover, 200 Monte-Carlo runs	Krippendorff's α 0.33 → 0.57; pass/fail α 0.34 → 0.63; pass/fail flip rate 57% → 31%	SIMULATED. The gain follows from the stated assumptions, mostly from the rubric anchors; the AI reveal changed about 5% of scores. Details
Consistency with real assessors	Live study mode is built (OFF → ON arms, α on the QA dashboard)	Not yet run with real assessors
Shadow rater (weld defect / quality level)	95% defect, 90% level accuracy on hold-out and on a fresh seed	SYNTHETIC images only. A demo model, not for certification. Details

Tests and CI

GitHub Actions (ci.yml) runs:

• ruff + 141 backend tests;
• eslint, vue-tsc, 119 unit tests and a production build;
• 17 Playwright e2e tests against the dev servers, plus an offline test against the built PWA with its real service worker;
• a Docker full-mode job that runs the compose stack and the smoke test.

Locally, Playwright drives an installed Chromium-based browser (PLAYWRIGHT_BROWSER_PATH) with a fresh profile and never downloads browsers.

npm test                              # backend + frontend unit tests and lint
cd frontend && npx playwright test    # e2e (starts the API and Vite itself)
npm run e2e:pwa --prefix frontend     # offline e2e against the built PWA
npm run screenshots --prefix frontend # regenerates docs/screenshots/

Honest limitations

• The shadow rater and the golden-packet consensus are synthetic. They show that the pipeline works, not how a model would do on real welds.
• The consistency gain is simulated. Real evidence needs the live study mode with real assessors.
• Voice input depends on the browser. It uses the browser's Web Speech API; Chrome on Android supports hi-IN, while Brave and Firefox usually don't, and typed text and icon cards are always available. The Bhashini adapter is a scaffold (HTTP 501), and offline speech (Vosk) is not built. Screenshot 1 uses a simulated recogniser so the voice UI can be shown in a headless browser.
• MySQL / SQL Server have not been run live. The migrations are compile-tested for both dialects, and setup_db.py is unit-tested.
• The Docker image omits torch, so full mode maps with BM25 + lexicon. pgvector is wired in and smoke-tested, but holds the embeddings only when they are enabled.
• The VC has not been checked with an external verifier. The proof follows eddsa-jcs-2022 with a did:key issuer. The SIDH export is a mock schema.
• pHash reuse detection can be evaded. Crops and rotations get past it, and about 1% of distinct synthetic welds collide (threshold 6, measured).
• One QP is official; ten are drafted for mapping breadth (marked "Demo QP" in the UI).

More detail, with the measurements behind each decision: docs/DECISIONS.md.

Docs

Demo script · API · Privacy (DPDP) · Architecture · Solution · Problem statement · Decisions · Environment

Licences

The project is under MIT (LICENSE). Every dependency is permissively licensed, checked against installed package metadata on 2026-10-20:

• Backend: FastAPI, Pydantic, SQLModel / SQLAlchemy, Alembic, PyJWT, pytest, ruff (MIT); uvicorn, scikit-learn, pandas, joblib, httpx, ImageHash, reportlab (BSD); bcrypt, python-multipart, rank-bm25, opencv-python-headless (Apache-2.0); cryptography (Apache-2.0 / BSD); numpy (BSD and compatible); Pillow (MIT-CMU).
• Optional embeddings: sentence-transformers, transformers, torch (Apache-2.0 and compatible). The model paraphrase-multilingual-MiniLM-L12-v2 is Apache-2.0.
• Full mode: pg8000, Celery, kombu (BSD); redis-py (MIT); minio-py (Apache-2.0). Images: PostgreSQL and pgvector (PostgreSQL Licence), SeaweedFS (Apache-2.0), Valkey (BSD-3), nginx (BSD-2).
• Frontend: Vue, Pinia, vue-router, vue-i18n, vue-echarts, qrcode, Vite, vite-plugin-pwa, Workbox, Tailwind CSS, Vitest, ESLint (MIT); Dexie, ECharts, Playwright, TypeScript (Apache-2.0).
• Not used, by licence:
  • MinIO server (AGPL-3.0) → SeaweedFS.
  • Redis 7.4+ (RSALv2/SSPL) → Valkey.
  • psycopg (LGPL) → pg8000.
  • The krippendorff package (GPL) → an in-house α, verified against it.
  • Ultralytics YOLO (AGPL).
• Disclosed:
  • certifi (MPL-2.0, an unmodified CA bundle) is pulled in transitively by httpx and minio.
  • matplotlib (PSF-style, dev-only) draws the result charts.

Team

SIH 2026 · PS 26242. Team name: PIXEL TITANS . Built by @Aribhussain ,@mdtabishgithub and Team.