MARK — Multi-Agent Research Kit

Overview

MARK (Multi-Agent Research Kit) is a deterministic equity-research pipeline: five domain-specialist agents — Price, Sentiment, Fundamentals, Macro, Risk — run in parallel through a single LangGraph fan-out/fan-in superstep. A supervisor QA gate checks each agent's output for missing data, contradictions, and short sections, and can force one retry round before synthesis. The pipeline assembles a sectioned markdown report with a labeled verdict, conviction level, and confidence score. A React + Vite single-page frontend is served alongside a FastAPI backend that wraps the pipeline.

What it does

• Five domain-specialist agents (Price, Sentiment, Fundamentals, Macro, Risk) run in parallel in a single LangGraph superstep, each with its own tool belt and data sources.
• A supervisor QA gate checks for missing data, contradictions, and short sections, and can request one retry round for flagged agents before forcing synthesis.
• A provider-agnostic data layer cascades per datatype (e.g. curated Supabase cache to LSEG/FMP/SEC EDGAR XBRL fallback) with LRU+TTL caching and capability-based routing.
• A RAG tool (semantic_search_filings) embeds the query and retrieves matched SEC-filing chunks plus full parent-section context from a Supabase pgvector store, scoped to a curated ticker universe.
• Synthesis assembles a sectioned markdown report — Executive Summary, Technical, Sentiment, Fundamentals, Macro, Risk, Synthesis & Verdict, Disclaimers — with a labeled verdict, conviction, and confidence score.
• Debate Mode lets users chat with each specialist agent post-verdict, stage refinements, and re-synthesize, with versioned thesis history and diffs.

Stack

Python · LangGraph · Anthropic Claude · Supabase · FastAPI · React + Vite

Status

Private venture, in active development — Debate Mode v1 and Supabase Auth (magic link + Google OAuth) are live.