We built Predict around the Moderna melanoma-vaccine story: the market-moving announcement was not an isolated event, but the visible end of a longer, evidence-backed sequence. The goal is to surface those connections earlier—before the story becomes obvious in the news.
- A React 19 + Vite dashboard for company records, company reports, and an interactive knowledge graph.
- An Express API that exposes companies, runs, graph data, reports, and health checks.
- A LangGraph workflow that runs Cala healthcare intelligence and source research in parallel, builds a relation pack, evaluates a healthcare gate, and only then performs the finance branch when the gate is positive.
- PostgreSQL with pgvector as the canonical store and Neo4j as a rebuildable graph projection.
- Source adapters for PubMed, ClinicalTrials.gov, RSS/news feeds, and Tavily web-news snippets.
flowchart LR
UI[React / Vite] --> API[Express API]
API -->|POST /runs| W[LangGraph worker]
W --> C[Cala healthcare]
W --> R[Research sources]
C --> J[Relation pack]
R --> PG[(PostgreSQL + pgvector)]
PG --> N[(Neo4j projection)]
N --> J
J --> G{Healthcare gate}
G -->|relevant and new| F[Cala finance + finance analysis]
G -->|otherwise| S[Persist completed run]
POST /runs queues work and returns immediately. The UI can poll the run and event endpoints while the worker progresses through fan-out, relations, healthcare gate, and (when applicable) finance.
- A current Node.js LTS release
- pnpm 9.15.0 (the version pinned in
package.json) - Docker Desktop, for PostgreSQL/pgvector and Neo4j
corepack enable
pnpm install
Copy-Item .env.example .env
pnpm db:up
pnpm db:migrate
pnpm devpnpm dev starts the API, scheduled worker, and frontend together. With the default configuration:
| Service | Address |
|---|---|
| Web dashboard | http://localhost:5173 |
| API | http://localhost:3002 |
| API health | http://localhost:3002/health |
| Neo4j Browser | http://localhost:7474 |
| PostgreSQL | localhost:15432 |
The dashboard is served at /; the interactive graph is at /knowledge-graph, and company details are at /companies/:companyId.
.env.example contains the local database defaults and optional provider variables. Add the keys needed for the data sources and model-backed analysis you want to enable:
OPENAI_API_KEY=
CALA_API_KEY=
TAVILY_API_KEY=
NEWS_FEED_URL=OPENAI_CHAT_MODEL defaults to gpt-5.6-luna; embeddings default to text-embedding-3-small. A missing Tavily key or RSS feed simply leaves that optional source empty.
pnpm dev # API, worker, and web app
pnpm dev:web # Vite dashboard only
pnpm dev:api # Express API only
pnpm dev:worker # Daily scheduler only
pnpm db:up # Start PostgreSQL and Neo4j
pnpm db:migrate # Apply Drizzle migrations
pnpm typecheck # Type-check all workspaces
pnpm test # Run workspace tests| Area | Endpoints |
|---|---|
| System | GET /health |
| Companies | GET/POST /companies, GET /companies/:id, timeline, people, developments, agent runs, and output sub-resources |
| Runs | POST /runs, GET /runs/:id, GET /runs/:id/events |
| Knowledge graph | GET /knowledge-graph, GET /knowledge-graph/entities/:id, POST /knowledge-graph/sql |
| Reports | GET /reports/momentum/:companyId |
The graph SQL endpoint generates and executes read-only queries; it never accepts Cypher from a client.