From AI prototype to production.

The gap between "it works in the demo" and "we can put this in front of people"

Plenty of small teams have built an AI assistant that works well most of the time — answering questions from documents, or turning referrals and forms into structured records. What they don't have is the apparatus around it: a privacy pre-step, an evaluation built from their own failures, and a gate that stops a quiet prompt tweak from shipping a system that fabricates. That apparatus usually requires an ML engineer. Groundwork scaffolds it instead.

Groundwork's units are deployment patterns — think Terraform modules for trustworthy AI. One command writes the whole harness for a pattern into your repo; one file connects it to your system. The checks are deterministic — same inputs, same result, every time, free.

See the gate catch fabrication before you wire anything

The repo ships fully offline demos — tiny systems wrapped in the exact harness init scaffolds. Run one, then flip the system into the failure modes real systems actually have and watch the build turn red:

That red build is the product: a fabricated figure, a guessed field, or a failed abstention becomes a CI failure, not a user complaint. The guessed date of birth is the failure worth staring at — the record looks more complete, and everything downstream trusts it.

Seven lessons, about two and a half hours

Every lesson ends with a checkpoint you can verify — not "you have read this," but "this specific thing now works on your machine." The biggest block is Lesson 3, because writing gold cases is the part that's actually yours. Take it on GitHub →

What a gold case looks like

Plain JSON, written by the person who knows the documents — no eval expertise required:

The scorer checks three things, deterministically: the anchored facts appear, every number traces to the context, and unanswerable questions get a refusal instead of an invention.

An extraction case is the same idea for records — the document, your JSON Schema, and gold labelled by hand, where null means the document does not state it:

Values compare exactly after per-field normalisation — dates in ISO, money in minor units — so there is no paraphrase problem and nothing for a judge model to judge. A value extracted into that null field fails, named per field.

Measure first, optimise second

groundwork cost reads real token usage from your eval runs and recommends savings in leverage order: prompt caching for the documents that ride along with every question, batching for anything not user-facing-urgent, context trimming (which is also safer), and model routing only after those — re-checked through the gate, because cheaper is only better if the grounding score doesn't move.

The same gate, conversationally

Groundwork ships an MCP server, so Claude Code, Cowork, and Claude Desktop can run the checks as tools — scaffold a harness for either archetype, run the full readiness gate on a repo, or paste a single answer (or a single extracted record with its schema and gold) and get the deterministic verdict with named issues, no repo required. The conversational checks import the exact same gate logic as CI, so the two can never drift apart.

There's also an Agent Skill that teaches an agent to run the gate and report results honestly — including refusing to present a green check as a safety certification. Both ship in the npm package.

Built on tools you can inspect