AI Agent Orchestration for Institutional Equity Research
Award: The Biggest Engineering Lift — Most Technically Involved compiled-7 / SanFranSokyo · YC RFS Spring 2026 AI Hackathon
Data Ingestion → Japanese NLP → Financial Modeling → Valuation Engine → Report Generation
(EDINET, TDNet, (Filings, (3-statement (DCF, comps, (Institutional-grade
IR sites, news) transcripts, models, scenario sum-of-parts) research output)
mgmt commentary) analysis)
| Layer | Technology |
|---|---|
| Frontend | Next.js 16, React 19, Tailwind CSS v4 |
| Backend | FastAPI, Python 3.12 |
| Agent Orchestration | LangGraph (sequential chaining & agent handoffs) |
| Adversarial Framework | AutoGen / AG2 (Bull vs. Bear debate loop) |
| Real-time Monitoring | Redis (100+ concurrent data streams) |
| LLM | Shisa.ai shisa-v2.1-llama3.3-70b (OpenAI-compatible API) |
| Storage | SQLite (decision traces), PostgreSQL (metadata), Vector DB (RAG) |
Key components:
- Swarm Topology — Specialized vertical agents (IR, Company, News, Satellite) each own their research domain
- Central Senior Analyst — Synthesizes sub-agent reports into an institutional-grade investment memo
- Adversarial Debate Loop — Bull Agent forms thesis → Bear Agent stress-tests with forensic accounting analysis
- Decision Traces — Every claim in the final report is clickable, linking back to the exact source (filing, transcript, audio)
- Omni-Channel Ingestion — Normalizes J-GAAP, IFRS, and US-GAAP data; monitors real-time macro feeds
cd backend
pip install fastapi uvicorn python-dotenv
# Optional: add SHISA_API_KEY to backend/.env for real LLM calls
uvicorn main:app --reload --port 8000cd frontend
npm install
npm run dev
# Open http://localhost:3000cd mcp
pip install -e .
python server.py