LiGHT Grant System

Dynamic Grant Intelligence, Tracking, and Proposal Automation Hub

An internal web platform for the LiGHT research group (EPFL) that manages the full grant lifecycle — from automated discovery and scoring through proposal development, archiving, and AI-assisted writing.

All AI workflows are powered by Qwen, running on your cluster.

---

Architecture

light-grant-system/
├── config.yaml            ← Master config (Qwen endpoint, scoring weights, schedule)
├── .env.example           ← Copy to .env and fill in secrets
├── docker-compose.yml     ← Full stack: DB, Redis, backend, workers, frontend
├── backend/               ← FastAPI (Python)
│   ├── app/
│   │   ├── ai/            ← Qwen client + all 10 agents + RAG retriever
│   │   ├── models/        ← SQLAlchemy ORM (17 tables)
│   │   ├── routers/       ← REST API endpoints
│   │   ├── scrapers/      ← Source connectors (RSS, HTML, API)
│   │   ├── services/      ← Business logic
│   │   └── workers/       ← Celery tasks (discovery, notifications, embeddings)
│   └── alembic/           ← Database migrations
├── frontend/              ← Next.js 14 (TypeScript + Tailwind)
└── scripts/               ← Seed data, admin creation

---

Quick Start

1. Configure Qwen

Edit config.yaml to point at your Qwen deployment:

ai:
  base_url: "http://your-cluster-host:8000/v1"  # vLLM / Ollama / TGI
  model: "Qwen/Qwen2.5-72B-Instruct"
  api_key: "EMPTY"  # or your auth key

2. Set environment variables

cp .env.example .env
# Edit .env — at minimum set SECRET_KEY

3. Start with Docker Compose

docker-compose up -d

This starts:

• PostgreSQL 16 with pgvector
• Redis
• FastAPI backend (port 8000)
• Celery worker + beat scheduler
• Next.js frontend (port 3000)

4. Run database migrations

docker-compose exec backend alembic upgrade head

5. Create admin user

docker-compose exec backend python scripts/create_admin.py admin@your-org.ch "Admin Name" yourpassword

6. Seed the 133 LiGHT opportunities

docker-compose exec backend python scripts/seed_opportunities.py /path/to/Opportunities.xlsx

7. Open the app

Visit http://localhost:3000 → log in with your admin credentials.

---

AI Workflows

All AI runs go through backend/app/ai/client.py, which calls the Qwen endpoint configured in config.yaml. Every run is logged in the ai_runs table with sources, outputs, and warnings.

Agent	Endpoint	Description
Call Analyzer	POST /api/v1/ai/analyze-call	Extracts structure from a grant call
Fit Scorer	POST /api/v1/ai/score-opportunity	Scores 0–100 against team priorities
Go/No-Go	POST /api/v1/ai/go-no-go	Strategic decision memo
Proposal Architect	POST /api/v1/ai/proposal-outline	Outline + timeline + assignments
Section Drafter	POST /api/v1/ai/draft-section	RAG-assisted section drafting
Compliance Checker	POST /api/v1/ai/compliance-check	Checks draft against funder requirements
Similar Grants	POST /api/v1/ai/find-similar-grants	Semantic search over archive
Feedback Analyzer	POST /api/v1/ai/analyze-feedback	Lessons from reviewer comments
Memory Agent	POST /api/v1/ai/process-for-memory	Archives completed grants

RAG / Vector Search

Proposal sections and documents are embedded using the Qwen embeddings endpoint (configured in config.yaml:ai.embeddings). Embeddings are stored in PostgreSQL via pgvector. The retriever uses hybrid search: vector similarity + keyword + metadata filters, with permission enforcement.

---

Configuration Reference (config.yaml)

Section	Key	Description
ai.base_url	Qwen server URL	vLLM/Ollama/TGI endpoint
ai.model	Model name	As registered on your server
ai.api_key	API key	"EMPTY" for local deployments
ai.embeddings	Embeddings config	Separate endpoint if needed
fit_scoring.team_themes	Theme list	Used by Qwen scorer
fit_scoring.tiers	Score thresholds	Customize tier boundaries
discovery.default_schedule	weekly/daily	How often to scan sources
rag.top_k	int	Results per retrieval query
notifications.email.enabled	bool	Enable email reminders

---

Development Setup (without Docker)

# Backend
cd backend
pip install -r requirements.txt
alembic upgrade head
uvicorn app.main:app --reload

# Celery worker (separate terminal)
celery -A app.workers.celery_app worker --loglevel=info

# Frontend
cd frontend
npm install
npm run dev

---

Grant Source Categories

Pre-configured source categories (add/edit via Admin UI or /api/v1/sources):

• Global health, AI & data science, Digital health
• Swiss funding (SNSF, Innosuisse, Swiss AI Initiative)
• EU funding (Horizon Europe, IHI, ERC, MSCA)
• US federal (NIH, USAID, FCDO)
• Private foundations (Gates, Wellcome, Grand Challenges)
• EPFL internal calls, Yale/Harvard institutional

---

API Documentation

Interactive docs available at:

• Swagger UI: http://localhost:8000/api/docs
• ReDoc: http://localhost:8000/api/redoc