FinSight Pro Connecting...
← Portfolio How it works Architecture & Pipeline
auto-detect
A Working Demonstration for Domain Research Agents

Researching a company used to take an afternoon. Now it takes one question.

Picture the work of researching a business: you bounce between a database of financials, a stack of hundred-page reports, and a spreadsheet running your own checks. It takes hours, and half the time you're just looking things up. FinSight Pro collapses all of that into a single question — it finds the numbers, reads the fine print, and flags anything risky, then writes back one clear answer with every source linked, in under 12 seconds.

The business problem

Finance analysts, compliance officers, and due-diligence teams spend hours a day reconciling numerical data with textual disclosures. A single question often requires 6–8 tool switches between a terminal, a filings reader, and a spreadsheet. That's the bottleneck.

Transferable across industries

The pattern — a routing agent orchestrating structured queries, document retrieval, and rule-based checks — generalizes to legal discovery, clinical decision support, insurance underwriting, vendor risk management, and regulatory intelligence. Same architecture, different knowledge base.

LangGraph Hybrid RAG Ragas Eval voyage-finance-2 Cohere Rerank 3
Architecture

Each step below is a LangGraph node — an AI agent with a specific job. LangGraph is the orchestration framework that decides which agents to call, in what order, based on your question.

QUERY
  └─▶ Router
        ├─▶ Data Agent
        │     └─▶ PostgreSQL
        │           FMP financials
        ├─▶ Doc Agent
        │     └─▶ pgvector hybrid
        │           BM25 + voyage-finance-2
        │           Cohere Rerank 3
        └─▶ Risk Agent
              └─▶ 6 arithmetic signals
                    └─▶ Synthesis → SSE
Pipeline
  1. Router (Gemini Flash) — classifies your question, identifies the company ticker, decides which agents to activate
  2. Data Agent — translates your question into SQL and queries 5 years of financial data (revenue, margins, FCF)
  3. Doc Agent — searches 532 SEC 10-K filing excerpts using hybrid BM25+vector search, reranked by Cohere
  4. Risk Agent — runs 6 arithmetic checks on the financials: liquidity, leverage, growth, coverage, margins, and cash flow
  5. Synthesis — Gemini Flash combines all findings into a cited answer and streams it live to your browser
Ragas Evaluation — P0 vs P1
Metric P0 Baseline P1 (Cohere Rerank) Delta
Faithfulness 0.606 0.683 +0.077
Context Precision 0.523 0.552 +0.029
Answer Relevancy 0.540 0.551 +0.011