Agentic Equity Researcher
Multi-agent AI system for automated, auditable equity research workflows
Overview
Agentic Equity Researcher is a production-grade AI system designed to automate large portions of the equity research lifecycle while preserving determinism, traceability, and analyst oversight. The system treats equity research as a structured workflow rather than a free-form chat interaction.
Agentic Architecture
- Designed a custom agentic orchestration framework coordinating multiple role-specific AI agents
- Implemented specialized agents for Data Ingestion, Financial Analysis, News & Filings, Risk & Consistency, and Report Synthesis
- Agents are executed through a DAG-based controller with explicit dependencies and deterministic execution guarantees
- All intermediate agent outputs are persisted, enabling reproducible re-runs and step-level inspection
LLM Infrastructure
- Built the system to run LLMs locally with a pure llama.cpp backend, avoiding heavyweight frameworks
- Tinkered extensively with llama.cpp model loading, GPU offloading, inference parameters, and embedding flows
- Implemented model routing and caching strategies to tightly control inference latency and compute cost
Retrieval & Advanced Citations
- Integrated Retrieval Augmented Generation (RAG) over structured financial data and unstructured documents including filings, transcripts, and research notes
- Implemented advanced citations for RAG responses, attaching document names, page numbers, highlighted text, and extracted images to each answer
- Built an embedded document reader allowing users to scroll cited documents inline and download annotated PDFs
- Designed the system to surface at least partial citations for every RAG-driven response where applicable so that the user can verify the source of the information instantly.
Performance & Impact
- Processed 10K+ documents across multiple equities using hybrid search and embedding caches
- Achieved sub-second retrieval latency for most research queries
- Reduced analyst turnaround time for coverage updates by approximately 60–70%
- Improved consistency and reproducibility of research outputs across repeated runs