CRIMSON
Intelligence-grade criminal network analysis.
CRIMSON turns raw investigative data into a living, queryable intelligence graph — giving analysts the ability to surface hidden criminal networks in minutes, not weeks.
The Story
Why this exists.
The Problem
Law enforcement analysts deal with thousands of unstructured data points — call records, witness statements, transaction logs — that are manually cross-referenced in spreadsheets. Connections are missed. Investigations stall.
Why It Matters
A single missed link in a criminal network can mean an investigation fails, a suspect walks free, or a network grows. Automated graph analysis with AI-assisted entity resolution changes the equation entirely.
The Solution
CRIMSON ingests structured and unstructured data, runs entity resolution to deduplicate identities, builds a persistent graph database, and exposes a query interface powered by Gemini AI. Analysts can ask natural-language questions and get graph-backed answers.
Product Features
What it does.
Live Criminal Network Graph
Interactive force-directed graph showing entities (persons, organizations, locations, events) and their weighted relationships. Zoom, filter, and cluster by connection strength.
AI Entity Resolution
Gemini AI extracts named entities from unstructured text — reports, transcripts, messages — and resolves duplicates using embedding similarity. A single person mentioned under five aliases becomes one node.
Natural Language Query
Analysts type plain English queries — 'Show everyone linked to X within 2 hops' — and the system translates to graph traversal queries, returning visual + tabular results.
Case Management
Organise entities and relationships into cases. Each case has an audit trail, analyst notes, and export to PDF/JSON for court submission.
Real-Time Ingestion
Webhook-based ingestion pipeline allows live data feeds from partner systems. New entities and edges appear in the graph within seconds.
System Architecture
How every layer connects.
Data flows: ingestion → AI extraction → graph storage → query layer → analyst UI
Technical Deep Dive
Under the hood.
Entity Resolution Pipeline
Raw text arrives via webhook or manual upload. The FastAPI service chunks the text, sends it to Gemini for NER, then generates sentence embeddings (text-embedding-004) for each extracted entity. A cosine similarity check against existing records (threshold 0.91) determines whether to create a new node or merge with an existing one. This reduces duplicate nodes by ~73% compared to string matching alone.
Graph Query Translation
Analysts write plain English. The system uses a structured prompt chain: (1) classify intent, (2) extract parameters (entity names, hop depth, relationship types), (3) generate a parameterised SQL query. Results are returned as both raw JSON and a subgraph for visualisation.
-- 2-hop network around an entity
WITH RECURSIVE network AS (
SELECT id, name, 0 AS depth
FROM entities WHERE name ILIKE $1
UNION ALL
SELECT e.id, e.name, n.depth + 1
FROM entities e
JOIN edges ed ON ed.target_id = e.id
JOIN network n ON n.id = ed.source_id
WHERE n.depth < $2
)
SELECT DISTINCT * FROM network;Security & Access Control
Every entity and case is scoped by organisation_id with Supabase Row Level Security. Analysts only see their organisation's data. All API calls are logged to an append-only audit table for full traceability.
Engineering Decisions
Challenge → Decision → Result.
Performance & Scale
By the numbers.
Deployment & Infrastructure
Production setup.
Deployment
Next.js on Vercel Edge with ISR. FastAPI on Render with a persistent worker dyno for the ingestion queue. PostgreSQL on Supabase with PgBouncer connection pooling.
CI/CD
GitHub Actions — lint + type-check on PR, Playwright smoke tests on merge, auto-deploy to Vercel and Render on green CI.
Monitoring
Vercel Analytics for frontend. Render metrics for API latency and memory. Supabase dashboard for DB query times and RLS audit.
Challenges & Failures
What broke — and how I fixed it.
What I Learned
Key takeaways.
- 01
Recursive CTEs in PostgreSQL can replace a dedicated graph DB for datasets under 1M nodes.
- 02
Embedding deduplication needs a human-reviewable confidence score, not just binary match/no-match.
- 03
NL → structured query is only reliable when you tightly constrain the output schema in the prompt.
- 04
Row Level Security in Supabase is more powerful than application-level filtering — push access control to the DB.
Future Roadmap
Where this is going.
v2.0 — Q4 2026
- —Temporal graph analysis — track network evolution over time
- —Multi-jurisdiction federation with privacy-preserving queries
- —Fine-tuned NER model on criminal investigation corpora
v3.0 — 2027
- —Mobile analyst app (React Native) for field use
- —OSINT API integration (Shodan, Maltego)
- —Anomaly detection — auto-flag unusual network patterns
Screenshots
The product.
Live criminal network graph — 847 entities
Entity extraction pipeline output
Natural language → graph traversal result
Technology Stack
Built with.
Frontend
Backend
Database
AI/ML
Auth / DB
Visualisation
Deployment
Ready to dive in?