VeriTrust

AI-powered medical misinformation detector. Users submit text or a URL; a LangGraph multi-agent pipeline (Extractor → Translator → Investigator → Health Expert) backed by Llama (via Ollama) and a fine-tuned BioBERT classifier returns a verdict (verdadera/falsa), confidence score, supporting biomedical sources, and medical explanation. The Investigator cross-checks each claim against Europe PMC literature, and the verdict's confidence is attenuated when no supporting evidence is found. Results are persisted in PostgreSQL and shown in a Next.js dashboard.

Stack

• Backend — FastAPI (web) + arq worker, LangGraph, Ollama, Transformers (BioBERT), Europe PMC evidence retrieval, psycopg3
• Frontend — Next.js 16 (App Router), React 19, Clerk, SWR, Tailwind v4
• Data — PostgreSQL, Redis (arq job queue)
• ML — BioBERT fine-tuned on PUBHEALTH

Analysis is asynchronous: POST /analysis returns immediately with a pending id, an arq worker runs the pipeline, and the frontend polls GET /analysis/{id} until the verdict is ready.

Repository layout

backend/
  app/        FastAPI service (routes, agents, db, schemas)
  ml/         BioBERT training and evaluation
  tests/      App test suite
frontend/
  src/        Next.js App Router

Prerequisites

• Python 3.11+ and uv
• Node.js 22+ and pnpm 9+
• Ollama running locally on :11434 with the models llama3, llama3.2, and translategemma pulled
• A PostgreSQL database (the Docker Compose stack provides one)
• A Redis instance on :6379 (job queue for the analysis worker)
• A Clerk application for auth

Onboarding

Quick start — Docker (recommended)

cp .env.example .env             # fill in Clerk values
cp backend/.env.example backend/.env
cp frontend/.env.example frontend/.env
docker compose up --build
docker compose exec ollama ollama pull llama3
docker compose exec ollama ollama pull llama3.2
docker compose exec ollama ollama pull translategemma

Frontend at http://localhost:3000, backend at http://localhost:8000. The BioBERT weights under backend/models/bert_classifier/ are mounted into the backend container read-only.

Local development (without Docker)

1. Configure environment

Copy the example env files and fill in real values:

cp backend/.env.example backend/.env
cp frontend/.env.example frontend/.env

Backend needs at minimum: DATABASE_URL, REDIS_URL, CLERK_JWKS_URL, CLERK_AUDIENCE, CORS_ALLOWED_ORIGINS, OLLAMA_BASE_URL. Frontend needs: NEXT_PUBLIC_API_URL, NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY, CLERK_SECRET_KEY.

2. Backend

cd backend
uv sync --frozen                      # install serving + api deps
uv sync --frozen --extra ml           # add the ML stack (only if training/evaluating)
uv run python -m app.main             # web API at http://localhost:8000
uv run python -m app.worker           # analysis worker (separate terminal; needs Redis + Ollama)

3. Frontend

cd frontend
pnpm install
pnpm dev                        # http://localhost:3000

4. Pull the Ollama models

ollama pull llama3
ollama pull llama3.2
ollama pull translategemma

5. BioBERT model

The detector loads weights from backend/models/bert_classifier/. Either copy a pre-trained checkpoint there or run the training pipeline:

cd backend
uv run python -m ml.training.train

Common commands

Backend (from backend/)

uv run pytest tests --cov=app --cov-fail-under=80     # app tests
uv run pytest ml/tests --cov=ml --cov-fail-under=80   # ML tests
uv run ruff check app ml tests                        # lint
uv run ruff format app ml tests                       # format
uv run mypy                                           # type-check

Frontend (from frontend/)

pnpm lint
pnpm test
pnpm build
pnpm generate:api-types         # regenerate src/types/api.d.ts (backend must be running)

Notes

• After changing a backend Pydantic schema, regenerate the frontend types with pnpm generate:api-types. Never edit frontend/src/types/api.d.ts by hand.
• The database schema lives in backend/db/init.sql, applied once to a fresh database (mounted into the Postgres container's docker-entrypoint-initdb.d). There is no migration framework yet — pre-launch, recreate the volume (docker compose down -v && docker compose up) to pick up schema changes.

License

MIT — see LICENSE.