ClearAgent, Agent Compliance & Audit Layer

Problem

The EU AI Act places obligations on autonomous agent systems — record-keeping, human oversight, traceability — that are written for organizations rather than for software. Turning them into something a system can actually enforce and be scored against is the hard part.

Approach

Three of those obligations are addressable in software, so the build was sequenced around them. A LangGraph planner-executor service coordinates 3 specialist subagents — verification, audit, and policy enforcement — across 2 LLM providers. Policy lookup pairs vector embedding search with cross-encoder reranking, so an agent action returns the one governing clause rather than a ranked list of plausible ones.

What it measured

3 scored dimensions define compliant behavior operationally: policy adherence, audit completeness, and decision provenance. A model-agnostic harness emits structured outputs against them, replacing manual review with deterministic, reproducible scoring.

Result

An event-sourced audit ledger with cryptographic chain-of-custody captures token usage, tool-call traces, and policy-violation flags on every agent invocation, which turns otherwise opaque agent behavior into data you can query. This is an engineering project rather than a product — it has no users.

On the engineer resume

• Built a LangGraph multi-agent orchestration service on a planner-executor pattern, coordinating 3 specialist subagents (verification, audit, policy enforcement) for autonomous tool-using AI agents across 2 LLM providers (OpenAI, Anthropic).
• Engineered a semantic retrieval pipeline pairing vector embedding search with cross-encoder reranking over the EU AI Act corpus, returning the single governing clause for an agent action instead of a ranked candidate list.
• Designed an event-sourced audit ledger with cryptographic chain-of-custody capturing token usage, tool-call traces, and policy-violation flags on every agent invocation, turning opaque agent behavior into queryable observability data.