AI agents that do real work, with a human holding the pen.

Anything that involves reading messy input and producing structured output or a draft at volume: qualifying replies, extracting fields from documents and speech, flagging risk in contracts, answering staff questions from your own data, writing first drafts of reports and questions. The six products above each replaced a manual job.

Three ways. Retrieval, so answers come from your data rather than the model’s memory. Structured outputs, so the model fills a schema instead of writing prose. And a split where the model extracts or proposes and deterministic code validates and decides. Plus logging and an evaluation set, so drift is caught.

Claude and OpenAI models for reasoning and extraction, Whisper for speech, chosen per job and swappable behind an interface you own. We build with Claude Code and Cursor day to day, so the tooling is not new to us.

No. We use the providers’ API terms that exclude training, scope what the model may see, redact what it does not need and set retention explicitly. Where the data is regulated we keep it inside the appropriate perimeter.

An agent wired into your tools with guardrails and logging ships in two to three weeks as a fixed-scope, fixed-price sprint, quoted within 48 hours of the scoping call. Price is driven by the number of tools and data sources, the interface and whether it sits inside a new product.

Agents drift as inputs change. Most clients keep a light retainer for prompt tuning against real cases, cost monitoring and new tools as the job expands. Everything stays in your accounts.