Risk Radar answers one question with engineering instead of vibes: is this Bitcoin address safe to transact with? It pulls the address's full history from five independent explorers, screens it against the U.S. Treasury OFAC SDN list and public abuse reports, extracts dozens of behavioral fingerprints from the raw transactions, and synthesizes everything into a letter grade with a complete evidence trail.

What makes it inventive is the conversation layer. Every assessment is exported as a structured, machine-readable knowledge graph — and that export is handed to a large language model. The result: you can literally chat with the blockchain. Ask “who likely controls this wallet?”, “why did this address fail?”, or “write me a compliance memo”, and the model reasons over verified on-chain facts rather than hallucinating.

The second inventive leap is inference from unknowns. The chain never tells you an address's owner, purpose, or intent — those are hidden variables. Risk Radar infers them from what is observable, using trusted-data analysis algorithms detailed below.

Knowledge graphsOFAC / sanctions screeningBehavioral feature engineeringMulti-source data fusionLLM report synthesisCo-spend clusteringProbability calibrationBTC-native billing

How it works

Fuse. The resolver queries five explorers in parallel (mempool.space, Blockstream, Blockchain.info, Blockchair, BlockCypher), reconciles disagreements, and merges a single trusted view of the address.

Screen. The merged identity is checked against the OFAC SDN sanctions list and Chainabuse-style public scam reports — the definitive, non-negotiable signals.

Fingerprint. A behavioral feature extractor computes peel-chain hops, fan-in/fan-out ratios, transaction velocity and burst windows, value irregularity (coefficient of variation), dusting, change-output ambiguity, consolidation, layering, round-amount USD inference, and circadian timing patterns.

Infer. Co-spend clustering groups addresses likely controlled by one actor; a seeded concept ontology (FATF, OFAC, FTC, FBI vocabularies) maps harvested web evidence onto fraud classes with source-trust weighting and multi-source consensus voting.

Score. Fifty-one dimensions roll up through six weighted sections (sanctions 30%, behavior 25%, evidence 15%, classification 15%, identity 10%, operational 5%) with a severity floor so a definitive sanctions hit can never be averaged away.

Converse. The full structured export is streamed to the LLM, which writes intelligence reports, answers follow-up questions, and generates downloadable PDFs — grounded in the graph, not guesswork.

FAQ

Which algorithms actually run under the hood?

Multi-source data fusion with per-field reconciliation; heuristic co-spend graph clustering; ratio and windowed-rate behavioral features; coefficient-of-variation dispersion analysis; source-trust weighted evidence scoring (government 0.95, sanctions 0.98, abuse databases 0.85); multi-source consensus voting calibrated into probabilities; a 51-dimension weighted linear scoring ensemble with max-composition severity flooring; three-band calibrated risk thresholding; and LLM-based natural-language synthesis over the structured export.

How can it know things the blockchain doesn't record?

It can't know — it infers. Ownership, purpose, and intent are hidden variables, but they leave observable shadows: timing rhythms suggest a timezone, co-spent inputs suggest common control, output patterns suggest exchanges or mixers, and web evidence anchors real-world context. Each inference carries an explicit confidence score and is labeled as inference, never presented as fact.

Is the ChatGPT layer just a chatbot bolted on?

No — that's the invention. The model never free-associates about an address. It receives the engine's complete RDF-style export: every score, signal, classification, and citation. It reasons over verified data, so its reports cite the same evidence trail the dashboard shows.

Is this financial or legal advice?

No. Risk Radar reports observable facts, public allegations, and model predictions — clearly labeled as each — so that you and your compliance process can make the decision.

Delivered as SaaS

Every product here ships the same way: software as a service. Nothing to install, nothing to patch — you open a URL and the current version is simply there. One codebase serves every customer (multi-tenant), data pipelines refresh continuously in the background, subscriptions replace licenses, and APIs let other software consume the same intelligence the dashboards show. Risk Radar demonstrates the full pattern: hosted assessment engine, tiered plans, Bitcoin-native invoicing, JSON-LD exports for machine consumption, and an LLM layer as a service on top of a service.