Transforming complex data into actionable intelligence. Our autonomous agent investigates across the web, mapping relationships and flagging risks in real-time.
6
Search phases
10
Pipeline stages
3+
AI models
100%
Real-time streaming
Launch a deep intelligence crawl.
How the Deep Research AI Agent works — from search and extraction to risk debate and report.
A simulated investigation of Timothy Overturf, CEO @ Sisu Capital
Director-driven loop (search → facts → risk → connections → verification) then synthesis: entity resolution, temporal analysis, Neo4j persist, graph reasoning (discovery queries), and report generation.
Live Metrics
Execution Log
1/16 entriesEverything you need for autonomous due diligence investigation.
Six distinct search phases — Baseline, Breadth, Depth, Adversarial, Triangulation, Synthesis — that loop adaptively until coverage is sufficient.
Risk Analyst and Devil's Advocate LLMs debate each risk finding, ensuring balanced assessments with mitigation factors.
Fuzzy-matching deduplication merges entities across aliases and co-references, building a clean entity graph.
Reconstructs chronological timelines, detects date inconsistencies, and surfaces career and association history.
Neo4j persistence with Cypher graph queries and interactive React Flow visualization for entities and relationships.
Watch the investigation unfold live — facts, entities, and risks appear as the agent discovers them via Server-Sent Events.
A 10-stage autonomous pipeline powered by LangGraph. The Director loops through stages 2–6 (Web Research, Fact Extraction, Risk Analysis, Connection Mapping, Source Verification) until coverage is sufficient; then stages 7–10 run once: Entity Resolution, Temporal Analysis, Neo4j persist, Graph Reasoning (discovery queries: centrality, paths, shell-company, shared-address), and Report Generation. Graph insights feed the report and Graph tab. Optional sign-in (Privy) and case persistence (Supabase) are available for deployment.
Orchestrates the investigation: chooses the next step (search, risk analysis, connection mapping, source verification, or generate report) and research phase based on coverage and diminishing returns.
Runs phased search queries (Baseline → Breadth → Depth → Adversarial → Triangulation) via Tavily (and Brave fallback) with result deduplication.
Extracts entities, connections, and facts with confidence scores and source URLs from retrieved content; batches by token budget.
Risk Analyst and Devil's Advocate LLMs debate to surface regulatory, reputational, financial, and legal flags with severity and mitigation.
Maps relationships between entities (e.g. WORKS_AT, BOARD_MEMBER_OF) with confidence; feeds the identity graph and Neo4j schema.
Cross-checks claims across sources and flags contradictions; improves confidence and supports risk and temporal analysis.
When the Director chooses report: deduplicates and merges entities via fuzzy matching and alias resolution for a clean graph.
Builds chronological temporal facts and detects contradictions (e.g. overlapping roles); feeds the Timeline tab and report.
Synthesizes the due diligence report from entities, connections, risks, temporal facts, and graph insights; supports PII redaction.
Persists entities, connections, and risk flags to Neo4j; runs graph discovery (degree centrality, shortest path subject→risk entities, shell-company detection) and appends insights to the report and Graph tab.
LangGraph
Agent orchestration
OpenAI
Deep analysis
Anthropic
Risk debate
Google Gemini
Report synthesis
Tavily
Web search API
FastAPI
SSE streaming backend
Next.js
React frontend
React Flow
Identity graph
Privy
Optional sign-in
Supabase
Case persistence