Solution
Elliptic extended its existing Data & Intelligence Platform with Agent Bricks to build Elliptic's copilot — an agent that describes wallet activity, highlights exposure paths, and produces neutral, report-ready risk narratives aligned to each client's risk appetite. The copilot is backed by Databricks Foundation Model APIs and Model Serving, and every interaction is captured by MLflow tracing so each prompt, intermediate step, and final narrative can be audited end-to-end. Automated evaluation pipelines score responses for correctness, relevance, and safety using LLMs-as-judges, with custom domain-specific safety scorers that distinguish describing money-laundering typologies (appropriate) from promoting harmful activity (inappropriate). Prompt and safety controls are versioned in MLflow so the team can iterate without rework.
Data flow
Wallet activity, transaction history, exposure values, screening results, and Elliptic's risk scores flow from the existing Data & Intelligence Platform into Agent Bricks, where Elliptic's copilot runs on models served via Databricks Foundation Model APIs and Model Serving. Every prompt, intermediate step, and final narrative is captured by MLflow tracing, with automated LLM-as-judge evaluation pipelines and custom domain-specific safety scorers feeding back into prompt and safety versioning. The resulting narratives are surfaced to compliance analysts at exchanges, banks, payment providers, and institutional custody clients.