README.md

---
title: DealFlow AI
emoji: 📊
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: mit
extra_gated_prompt: Please provide your HF_TOKEN and SERPER_API_KEY as Space secrets.
---

# DealFlow AI

**Multi-Agent Investment Due Diligence System**
AMD Developer Hackathon 2026 | CrewAI v1.0 + Qwen2.5-72B + HuggingFace Inference

---

## What It Does

Upload a startup pitch deck PDF → DealFlow AI deploys a 3-agent crew that:

1. **Researcher** — extracts key claims from the pitch deck, searches the web for market validation and competitor analysis
2. **Financial Analyst** — validates financial projections, models unit economics using Python/pandas
3. **Report Writer** — synthesizes all findings into a professional investment memo

## Quick Start (Local)

```bash
# 1. Enter project directory
cd dealflow-ai

# 2. Copy and fill in env vars
cp .env.example .env
# Edit .env: SERPER_API_KEY, HF_TOKEN

# 3. Install dependencies
pip install -r requirements.txt

# 4. Launch the app
python app.py
# → Opens at http://localhost:7860
```

## Agent Tools (per PRE-FLIGHT SPEC #1 — exact, no improvisation)

| Agent | CrewAI Tools |
|-------|-------------|
| Researcher | SerperDevTool, ScrapeWebsiteTool, WebsiteSearchTool, PDFSearchTool, FileReadTool |
| Financial Analyst | CodeInterpreterTool, CSVSearchTool, JSONSearchTool, FileReadTool |
| Report Writer | FileReadTool, RagTool |

## Inference Stack

| Component | Value |
|-----------|-------|
| Model | `Qwen/Qwen2.5-72B-Instruct` |
| Provider | HuggingFace Serverless (featherless-ai router) |
| Search | Serper.dev API |

## Deployment

### AMD Fallback Protocol (PRE-FLIGHT SPEC #3) — **ACTIVATED**

| Checkpoint | Condition | Action |
|-----------|-----------|--------|
| T+12h | MI300X not provisioned | Begin parallel HF Endpoints setup as hot standby |
| **T+60h** | Still no AMD credits | **LOCK HF fallback. Abandon AMD-specific bonus.** |
| T+100h | MI300X recovered late | Do NOT switch back. HF is locked. |

> **AMD Fallback ACTIVATED at T+63h (2026-05-06).**
> AMD MI300X credits did not arrive by the T+60h checkpoint.
> Primary backend switched to HuggingFace Serverless Inference API (`Qwen/Qwen2.5-72B-Instruct` via featherless-ai provider).
> The AMD MI300X hardware-specific judging criterion is forfeited.
> The project qualifies for the **AI Agents** track and **HuggingFace Space** bonus criterion.
> Pitch narrative: "AMD-optimized architecture, deployed on HuggingFace infrastructure."

### HF Space Environment Variables

Set these as Space Secrets in the HF Space settings:

| Variable | Description |
|----------|-------------|
| `SERPER_API_KEY` | Serper.dev API key for web search |
| `HF_TOKEN` | HuggingFace token for inference |
| `LLM_BACKEND` | Set to `hf` (default) |

## Configuration

See `.env.example` for all options.

## Running Tests

```bash
pip install pytest
pytest tests/ -v
# Expected: 19 passed, 1 skipped
```

## Architecture

See [03-mvp/architecture.md](03-mvp/architecture.md) for full architecture diagram.

---

*Built for AMD Developer Hackathon 2026*
*DealFlow AI — Multi-agent investment due diligence on HuggingFace Spaces*