Experiments
Labs: agentic AI experiments and the patterns behind them
Working ideas I build to test new agent patterns: voice, retrieval, tools over the Model Context Protocol, evals, research crews and document extraction. Each one has its architecture, a simulated run and its stack.
The experiments
Voice receptionist
A voice agent that answers the phone for a business. It transcribes the caller in real time, works out what they want, finds and books a free slot in the calendar, answers in a natural voice and hands the call to a person when it should not decide alone.
Knowledge assistant
An assistant that answers questions from a company's own documents and cites the passage behind every claim. Each question is rewritten into a better search, matched against the embedded documents, reranked by relevance and answered with citations.
CRM tool server
A Model Context Protocol server that gives Claude a small set of safe tools for a CRM: search contacts, open tickets, move deals and summarise accounts. Every call passes a permission check and writes an entry to an audit log.
Evals and tracing
An evaluation and tracing setup that treats agent quality like a test suite. Every run is traced step by step, answers are scored by an LLM judge against a rubric, saved cases are replayed before each release and the cost of every run is tracked.
Research crew
A team of agents that researches a market question and writes the brief. A lead agent splits the question, a searcher collects sources, a reader summarises each one, an analyst compares claims and flags conflicts, and a writer delivers a two-page brief with its sources.
Document intelligence
An extraction pipeline that turns invoices and contracts into clean, validated data. Documents are classified by type, the fields that matter are extracted, totals, dates and formats are validated, and anything uncertain goes to a person before the rows are exported to the ERP.