Investment research, structured into a terminal.
Klypup is a multi-tenant investment research dashboard. An analyst types a natural-language query — an agentic LLM flow orchestrates market data, news sentiment analysis, and SEC filing vector search — and the results render as structured UI with cited sources, comparison tables, and charts.
Built as a three-tier architecture: React/Vite SPA (Vercel) → FastAPI backend (Render) → Supabase Postgres + pgvector (DB, auth, vector store).
- Natural-language research — Ask "Compare NVDA and AMD: revenue, valuation, and risks" and get a structured, sourced desk note
- Agentic AI flow — Router LLM picks which tools to run (market, news, filings); synthesizer LLM writes the analysis; editor LLM self-critiques
- Live market tape — Scrolling ticker with real yfinance data, cosmetic price jitter, and flash animations on tick direction changes
- SEC filing vector search — pgvector similarity search over 10-K/10-Q passages with Gemini embeddings and source citations
- News sentiment — DuckDuckGo news with lexicon-based positive/negative/neutral scoring
- Multi-tenant with RBAC — Org-scoped data isolation enforced in app code; admin/analyst roles
- Dark/light themes — OKLCH design tokens, terminal aesthetic, reduced-motion support
- ⌘K command palette — Quick navigation, theme toggle, and direct research queries
docker-compose upFrontend at http://localhost:5173, backend at http://localhost:8000.
# Frontend
cd frontend && npm install && npm run dev
# Backend (separate terminal)
cd backend && python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt && uvicorn app.main:app --reloadSee CONTRIBUTING.md for the full setup guide (database migrations, seeding, env vars).
Email: admin@acme.test
Password: demo1234
| Guide | Description |
|---|---|
| Architecture Guide | System diagrams, data flow, AI orchestration, embedding pipeline, tenant isolation |
| API Reference | All endpoints, request/response schemas, error codes |
| Decision Log | Architecture Decision Records and trade-off rationale |
| Contributing Guide | Setup, workflow, conventions, CI/CD |
User query → Router LLM → parallel tools → Synthesizer LLM → structured UI
│ │
picks tools yfinance, DDG news,
& tickers pgvector filings
The backend runs two LLM calls. The router turns a natural-language query into a JSON execution plan. The backend runs only the selected tools concurrently (asyncio.gather). The synthesizer produces a strict JSON UI state with every data point carrying a source citation. A third editor LLM self-critiques the draft and triggers a refine pass if gaps are found.
| Layer | Technology |
|---|---|
| Frontend | React 19, TypeScript, Vite 5, Tailwind CSS, Recharts |
| Backend | Python 3.12+, FastAPI, Uvicorn |
| Database | Supabase Postgres + pgvector |
| Auth | Supabase Auth (ES256 JWTs) |
| LLM | Google Gemini 2.5 Flash (primary), Groq Llama 3.3 70B (fallback) |
| Market data | yfinance |
| News | DuckDuckGo News |
| Embeddings | Gemini text-embedding-001 (3072-dim) |
| Vector search | pgvector IVFFlat, cosine distance |
| CI/CD | GitHub Actions (ruff, pytest, vitest, tsc, gitleaks) |
Klypup is a 5-day take-home assessment. All phases are substantially built. The frontend has undergone a visual redesign (Bloomberg-modern terminal aesthetic). See plan.md for the original build plan and docs/DECISIONS.md for trade-off rationale.