I built Tallybook, an offline‑capable multi‑agent stock research tool hosted on my GitHub. It uses a React + Vite front‑end, FastAPI server on port 8030, and a LangGraph engine to coordinate agents.
The architecture is "React + Vite site --> FastAPI :8030 --> LangGraph".
Agent 1 fundamentals (valuation, margins, growth, ROE, leverage) weight 35%.
Decision thresholds: Buy at +0.25, Sell at -0.25, otherwise Hold.
Each agent is wrapped. A failure becomes status: degraded with the reason, weights renormalise over the rest, the report still renders.
Test suite runs with "python -m pytest -q # 32 tests, no network, no model".
I provide CLI commands, Docker compose, and configuration instructions so anyone can run Tallybook offline or with live data, and the LLM only generates the final prose brief.