ProbashiNest — Probashi Shield

Probashi Shield checks a Bangladeshi migrant worker's job offer before they pay a recruiting agent. The worker speaks in Bangla or photographs the offer letter, visa or contract, and gets a red / amber / green verdict with reasons and next steps, spoken back in Bangla.

Built for the International AI Builders Congress 2026 (Idea to Unicorn, JusticeSphere AI).

How it works

Worker (voice or photo)
  → probashi.voice.transcribe_bn      Bangla speech to text
  → probashi.extract.extract_offer    OCR + key facts (employer, salary, agency, fees, country)
  → probashi.knowledge.evaluate       rules first, then RAG with cited sources → verdict
  → probashi.voice.synthesize_bn      spoken Bangla answer
  → frontend                          traffic-light result, reasons, next steps, helplines

Repository layout

Folder	Owner	What it holds
probashi/schemas.py	Shared	Data contract used by every module (do not edit)
probashi/llm.py	Shared	Swappable LLM wrapper
probashi/extract/	Agent 3	Document reading and fact extraction
probashi/knowledge/, data/	Agent 2	Knowledge base, verdict engine, test cases
probashi/voice/	Agent 4	Bangla speech-to-text and text-to-speech
backend/, infra/	Agent 1	FastAPI API, integration, AWS deployment
frontend/	Agent 5	Mobile-first web app

Rules for contributors (human or AI) are in CLAUDE.md.

Run locally with stubs

No API keys are needed. Stubs stand in for the extract, knowledge and voice modules and return synthetic, always-amber answers.

python3.11 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env          # USE_STUBS=true is already set there
pytest
uvicorn backend.main:app --reload

Then try it:

curl localhost:8000/health
curl -X POST localhost:8000/api/check -F 'text=সৌদি আরবে ক্লিনারের চাকরি'

Interactive API docs are at http://localhost:8000/docs. Or run the backend and the frontend together with Docker:

docker compose up --build     # backend on :8000, frontend on :5173

Switch to the real modules

Each module folder starts with a placeholder that raises NotImplementedError; the owning agent replaces it.

1. In .env, set USE_STUBS=false and fill in ANTHROPIC_API_KEY and the voice provider settings.
2. Restart the backend. GET /health should show "stubs": [].

While only some modules are ready, stub the rest by name, for example USE_STUBS=voice or USE_STUBS=voice,knowledge. A module that is neither implemented nor stubbed makes /api/check return 503 MODULE_NOT_READY.

Any response built with a stub carries the header X-Probashi-Stubs. If USE_STUBS is not set at all, the backend uses the real modules, so a deployment cannot serve fake verdicts by accident.

More

• API reference, error codes and privacy notes: backend/README.md
• AWS deployment, env vars and cost estimate: infra/README.md