MarketScout AI

Autonomous multi-agent market intelligence platform — powered by AMD Instinct GPUs via Fireworks AI, orchestrated with LangGraph.

Built for the AMD Developer Hackathon: Act II — Unicorn Track.

📖 Technical Report — full documentation of how the system works (architecture, agent pipeline, scoring methodology, evaluation).

---

Run It in One Command

No setup, no .env editing, no docker-compose. One self-contained image with everything baked in:

docker run -p 5000:5000 gam5510/marketscout-allinone:latest

Then open http://localhost:5000 — the app is ready immediately (demo mode, no login required).

---

What it does

You describe a startup idea. MarketScout AI deploys 15 specialised AI agents in a stateful LangGraph pipeline to research every dimension of the market:

Step	Agent	Output
0	Idea Guard	Validates legality, ethics, feasibility
1	Research Agent	Market overview, TAM, pain points
2	Competitor Agent	Competitor profiles, saturation score
3	Scientific Agent	Literature review, research maturity
4	Patent Agent	IP landscape, freedom-to-operate
5	Funding Agent	VC activity, funding rounds
6	Trend Agent	Market trends, adoption signals
7	Research Gap Agent	Novelty score, white spaces
8	SWOT Agent	Full SWOT matrix
9	Opportunity Agent	Scored opportunity list
10	Risk Agent	Risk registry with mitigations
11	Innovation Scoring	Composite score 0–100 with grade
12	Validation Agent	Go/no-go verdict
13	Strategy Agent	Go-to-market plan
14	Report Generator	Full narrative report

On completion the platform generates on-demand:

• 📄 PDF report — formatted market intelligence document
• 🎤 Pitch deck — 10-slide investor deck
• 💼 Business plan — full investor-ready business plan
• 🔭 Scenario simulation — what-if analysis without re-running the pipeline
• ↔️ Idea comparison — deterministic side-by-side comparison of two research jobs
• 🧭 Knowledge graph — interactive evidence network (vis.js)

All LLM inference runs on AMD Instinct MI300X GPUs via Fireworks AI.

---

Architecture

browser
  │
  └── nginx :5000
        ├── /api/agents/* → agent-service  (FastAPI + LangGraph)
        │                        ├── Fireworks AI (AMD Instinct GPUs)
        │                        └── Tavily Search API
        ├── /api/*         → backend       (FastAPI + Supabase)
        └── /*             → Next.js frontend

Layer	Technology
Frontend	Next.js 13 (App Router), TypeScript, Tailwind CSS, Radix UI, Recharts
Backend	FastAPI, Supabase/Postgres, Authlib (GitHub + Google OAuth), JWT cookies
Agent Service	FastAPI, LangGraph, 15 AI agents, Fireworks AI, Tavily
Infrastructure	Docker, nginx

---

LLM Models (all on AMD Instinct GPUs via Fireworks AI)

Model	Use Case
DeepSeek V4 Flash	Fast research agents (default)
DeepSeek V4 Pro	Strategy, validation, reasoning
Qwen V3.7 Plus	SWOT synthesis
MiniMax M3	Long-context report generation
GPT-OSS 20B	Lightweight tasks
GPT-OSS 120B	Heavy synthesis

---

Security

• 7-layer defence-in-depth (input validation, UUID injection prevention, rate limiting, JWT auth, job ownership isolation, secret scanner, prompt injection guard)
• Non-root Docker users in all three containers
• All .env files excluded from git

---

License

MIT