Multi-agent stock analysis pipeline

Three-agent system (analyst → challenger → synthesizer) running on locally hosted Gemma via Ollama, with a Streamlit dashboard.

Personal project · 2026

Idea

Single-LLM “stock analyst” demos tend to produce confidently wrong analysis because the model never has to defend its own claims. This project explores a three-role architecture where roles are deliberately adversarial:

• Analyst — surfaces the bullish or neutral case from available data.
• Challenger — argues the bearish case, identifies weaknesses in the analyst’s reasoning, surfaces disconfirming evidence.
• Synthesizer — adjudicates the disagreement and produces a final reasoned summary, including stated uncertainty.

The structure forces explicit disagreement before synthesis, which surfaces edge cases that a single-agent pipeline would miss.

Implementation

• Inference — locally hosted Gemma via Ollama on an RTX 4090 Linux desktop. Migrated to Gemini 2.5 Flash via the Google AI Studio free tier for the public demo (in-app rate limiting protects the quota).
• Frontend — Streamlit dashboard that exposes each agent’s reasoning side-by-side, so the disagreement is legible to the user.
• Optimization — careful inference and memory tuning across Apple Silicon (MLX) and Linux GPU environments.

Demo

Live demo on Hugging Face Spaces. A pre-recorded walkthrough video is available as a fallback if the demo is rate-limited.

Live demo →

Stack

Python, Ollama, Gemma, Streamlit, Google Generative AI SDK.