FinAgent: AI Trading Research on AMD MI300X

Retail traders face a daily triage problem: which of these ten tickers deserves attention today? Most AI tools answer with a single-shot prompt into a general-purpose model, confident, shallow, indistinguishable across tickers. Real analysts don't work that way. They divide the labour: one reads the news, another digs into financials, another reads charts, another sizes the trade, then a head of strategy synthesises. FinAgent reproduces that division with five CrewAI agents running against ONE locally-hosted Qwen3-14B instance on an AMD Instinct MI300X GPU. Market Scanner, Fundamental Analyst, and Technical Analyst run in parallel; Risk Manager consumes technical output to compute ATR-based position sizing; Chief Strategist synthesises the final signal. Every output is grounded in real yfinance quotes, with sanity guards that clamp runaway stops and re-ground the signal if the LLM drifts from the live price. The entire pipeline is self-hosted inference, no OpenAI, no Claude, no API bills. vLLM 0.17 on ROCm 7.2 exposes an OpenAI-compatible endpoint; CrewAI agents talk to it through the `hosted_vllm/` litellm provider with native tool-calling. The $100 AMD Developer Cloud credit is the entire compute budget. The frontend is a dark-themed Gradio 5 Space under the hackathon org, with a live agent-activity feed driven by structured callback events. Ten keyless tool functions wrap yfinance, DuckDuckGo news (via ddgs), and pandas-ta — no API keys anywhere in the demo. 309 unit + property-based tests (Hypothesis) pass on every commit, verifying correctness invariants end-to-end. Outcome: institutional-grade research, retail-friendly pricing, 100% open source.