AI Multi-Agent Stock Analyzer
Domains
Tech Stack
Project Summary
Abstract
AI Multi-Agent Stock Analyzer is a privacy-first financial research assistant built to run entirely on local infrastructure. The project uses Ollama-hosted language models and a Retrieval-Augmented Generation pipeline so stock analysis can be grounded in private reports, earnings transcripts, and custom research documents instead of depending only on model memory.
The core architecture is organized around a router agent that interprets user prompts and hands the work to specialized analyst agents. That division of labor makes the system easier to reason about and lets each sub-agent focus on a distinct slice of the problem, including technical analysis, fundamentals, sentiment, corporate events, and earnings interpretation.
To keep the workflow reproducible and easy to run, the full stack is containerized and exposed through a Streamlit interface. The result is a practical example of how local LLMs, document retrieval, and multi-agent orchestration can be combined into a usable financial-analysis product without sacrificing privacy.
What I Built
- A router-led multi-agent setup made it easier to decompose investor questions across technical, fundamental, sentiment, events, and earnings specialists.
- Local RAG over private documents reduced reliance on model memory and created more grounded analysis paths.
- Running Ollama models locally preserved privacy without sacrificing a usable interactive workflow.
Impact
- Delivered a practical private-by-default financial research assistant rather than a generic chatbot demo.
- Shows strong AI systems thinking across orchestration, retrieval, deployment, and product packaging.
Page Info
Router-Led Multi-Agent Workflow
Built a router agent that inspects each investor question, drafts an execution plan, and dispatches the task to specialized agents for technicals, fundamentals, sentiment, events, and earnings analysis.
Local RAG Over Private Financial Documents
Ingested local TXT, PDF, Markdown, and HTML documents into a vector store so the system retrieves grounded evidence and returns analysis with source-aware context instead of relying on raw model memory.
Private-by-Default Deployment
Wrapped the app in Streamlit and Docker Compose so the interface, vector pipeline, and Ollama-served models can run fully on local infrastructure without leaking proprietary research to external APIs.