Trader Advisor — agentic stock research
Personal project · 2026
Idea
Most LLM “stock analyst” demos are single-call: dump the data, ask for a verdict. The model never has to decide what data to look at, because everything is pre-fetched. This project flips that — an agent that picks its own tools during the investigation, calling for price context, indicators, options, sentiment, or fundamentals based on what it has already learned. The agent’s reasoning trace is the artifact, not just the verdict.
Three classical LLM pipelines (solo, core, full) ship alongside the agent as additional voices, so the same ticker can be looked at four ways.
Implementation
- Agent loop — hand-rolled tool-calling loop, no LangChain or LangGraph. Persistent message history, ephemeral system prompt rebuilt each turn with the current call budget, force-finalize on budget exhaust via
tool_choice="none". Eleven tools span price action, technical indicators, fundamentals, options flow, insider transactions, news, sector comparisons, and social sentiment. - Other modes — solo is one LLM call with all data dumped in (baseline). Core is a three-step adversarial workflow (analyst → challenger → synthesizer); flips ~30% of solo’s verdicts. Full wraps the open-source TradingAgents framework for a seven-agent debate.
- Inference — primary path is locally hosted gpt-oss:20b via Ollama on a single RTX 3090 (24GB). Provider-pluggable: provider abstraction designed for Anthropic, OpenAI, and Google Generative AI; primary deployed path is Ollama.
- Frontend — Streamlit dashboard with three tabs (Overview, Run Explorer, About) and a modal-driven run configurator. The agent trace renderer shows each step’s reasoning, the tool called, and the truncated tool result — legible as a journal of the investigation.
- Persistence — every run journaled to SQLite: prompts, tool calls, token cost, runtime, decision rationale.
Demo
Live demo on Hugging Face Spaces →
The deployed demo runs in read-only mode (pre-loaded analyses, live runs disabled because the agent needs Finnhub, Alpha Vantage, and an LLM key). The Run Explorer tab and agent trace viewer are fully interactive over the pre-loaded data.
Stack
Python, Ollama, Streamlit, SQLite, Plotly. Provider abstraction designed for Anthropic, OpenAI, and Google Generative AI; primary deployed path is Ollama.