AI-powered assistant for faster IFRC disaster response applications
DREF Assist helps emergency surveyors complete DREF (Disaster Response Emergency Fund) applications faster and to a higher standard. It accepts multimodal inputs — text, PDFs, images, voice recordings, and video — processes them through Azure OpenAI GPT-4o, and automatically populates the DREF application form via a conversational chat interface with real-time SSE streaming. When conflicting data is detected across sources, the system pauses for human resolution. A built-in evaluation engine scores the completed application against IFRC rubric criteria and suggests improvements.
The system is stateless — the frontend sends the full form state, conversation history, and files with every request. No database or session storage is required. The
/api/chatendpoint uses SSE (Server-Sent Events) to stream responses in real time, while/api/evaluatereturns standard JSON.
- Multimodal input — upload PDFs, DOCX, images, audio recordings, and video (one file per message); each processed by a specialised handler
- Real-time streaming — SSE-based streaming delivers AI responses token-by-token for a smooth conversational experience
- Auto field population — GPT-4o extracts structured data from uploads and chat messages, mapping it to 56 DREF form fields
- Conflict resolution — detects contradictions between new data and existing form values, pauses for user decision
- Automated evaluation — two-pass scoring (rule-based + LLM) against the IFRC rubric with actionable improvement suggestions
- IFRC GO integration — UI styled to match the existing IFRC GO platform to avoid retraining field officers
- Conversational interface — natural-language chat with full conversation history for iterative form completion
- Multilingual support — GPT-4o handles input in any language; Whisper transcribes non-English audio
DREF-Assist/
├── README.md ← You are here
├── backend/ ← FastAPI backend — AI pipeline, evaluation, API
│ ├── app.py Entry point & route definitions
│ ├── services/ Orchestrator (media → LLM → conflicts)
│ ├── llm_handler/ GPT-4o prompt building, calling, parsing
│ ├── media-processor/ Handlers for PDF, DOCX, image, audio, video
│ ├── conflict_resolver/ Conflict detection & resolution logic
│ ├── dref_evaluation/ Rubric-based evaluation engine (43 criteria)
│ ├── tests/ End-to-end tests
│ └── README.md Detailed backend docs
├── prototypefrontend/ ← React frontend — DREF form UI + chat + evaluation
│ └── drefprototype/
│ ├── src/
│ │ ├── pages/ Multi-step form wizard
│ │ ├── components/ Chat, evaluation panel, form steps, IFRC header/footer
│ │ └── lib/ API client, type definitions, utilities
│ └── README.md Detailed frontend docs
└── docs/ ← Design documents & plans
| Component | Technology |
|---|---|
| Backend framework | FastAPI (Python 3.10+), Uvicorn |
| LLM | Azure OpenAI GPT-4o (JSON mode, temperature 0.1) |
| Speech-to-text | Azure OpenAI Whisper |
| Media processing | PyMuPDF (PDF), python-docx (DOCX), OpenCV (video frames), Pillow, imagehash |
| Frontend framework | React 18 + TypeScript, Vite 7 |
| UI components | shadcn/ui (45+ Radix primitives), Tailwind CSS 3.4 |
| Forms | React Hook Form + Zod validation |
| Server state | TanStack React Query |
| Streaming | SSE (Server-Sent Events) for real-time chat responses |
| State architecture | Stateless — enriched form state (value + source + timestamp) sent per request |
- Python 3.10+ and pip
- Node.js 18+ and npm
- FFmpeg (for video frame extraction) —
brew install ffmpeg/apt install ffmpeg - Azure OpenAI resource with GPT-4o and Whisper deployments (setup guide)
git clone https://github.com/IFRCGo/DREF-Assist.git
cd DREF-Assistcd backend
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Create .env with your Azure OpenAI keys (see backend/README.md for all variables)
uvicorn app:app --reload --port 8000See backend/README.md for the full environment variable table and detailed setup.
cd prototypefrontend/drefprototype
npm install
npm run devOpen http://localhost:8080 — you should see the IFRC GO-styled DREF form with the AI chat assistant in the bottom-right corner.
See prototypefrontend/README.md for detailed frontend configuration.
| Resource | URL |
|---|---|
| Backend documentation | backend/README.md |
| Frontend documentation | prototypefrontend/README.md |
| UCL IXN Programme | ucl.ac.uk/computer-science/collaborate/ucl-industry-exchange-network-ucl-ixn |
| IFRC DREF Information | ifrc.org/disaster-response-emergency-fund-dref |
| IFRC GO Platform | go.ifrc.org |
DREF Assist is a UCL COMP0016 IXN project (2025–26) built for the International Federation of Red Cross and Red Crescent Societies (IFRC).
| Name | GitHub |
|---|---|
| Fahad Al Saud | @fbs617 |
| Brendan Loo | @brendanlhm |
| Sameer Chowdhury | @1sameerchowdhury |
| Mohammed Talab | @MohiCodeHub |
| Emir Akdag | @emirakdag0 |
Client: IFRC — International Federation of Red Cross and Red Crescent Societies Academic context: UCL Department of Computer Science — COMP0016 Systems Engineering (IXN), 2025–26