Case study / Product build
An AI copy editor that deploys into your cloud, not a vendor's
Newsrooms need house-style enforcement and cited fact-checking at deadline speed, and increasingly need the AI to run inside their own infrastructure. StyleProof is a copy-desk tool built around exactly that constraint.
Working build, verified in a customer's cloud- Host-agnostic engine
- Eval gate in CI
- Verified in a customer's Bedrock
The engagement
Challenge
Two mechanical jobs, under deadline, on someone else's cloud
Enforce a house style guide and check that a story's claims hold up: both slow, both error-prone. And media companies increasingly want the AI to run inside their own cloud, not a vendor's, which most tools are not built to do.
Action
A portable engine with a quality gate
A host-agnostic verification engine with zero HTTP or UI knowledge (enforced by a lint rule, not a convention), models and search behind swappable interfaces, a tiered agentic fact-checker, and a streaming review UI where a human accepts or rejects each item. Quality is gated in CI against a committed eval baseline.
Result
The portability claim stopped being a design goal
The same build was containerized and stood up inside a media customer's isolated AWS, on their Bedrock through an IAM role with no API keys, behind their SSO. Verified there end to end: it contradicted a planted false claim with citations, and when a real production symptom appeared (interrupted streams on slow fact-check runs) the fix was measured and re-verified under load. Clean-room, so it is mine to show and to sell.
How it is built
The shape that makes it portable
Models and search sit behind interfaces, so the same engine runs on a demo host or in a customer's isolated account.
- 01Engine purity. A lint rule fails the build if anything in the engine imports a framework or transport.
- 02Model map. Per-task models resolved through a ModelMap with no model ids in code at all; each deployment supplies its own five task models as configuration, all through Bedrock.
- 03Search interface. Tavily behind a SearchProvider, swappable per tenant.
- 04Skeptic panel. High-consequence verdicts get a second adversarial pass: three skeptic lenses that can escalate a claim to needs-human review but can never upgrade a verdict or invent confidence.
- 05Eval gate. npm run eval fails on F1 or fact-status accuracy dropping more than 5 points, cost per article rising more than 25 percent, or any blacklisted false positive.
- 06Reviewed baselines. Regenerating the baseline prints an old-to-new diff that is the review artifact in the pull request, so a ratified regression is never silently absorbed.
What it proves
The eval gate answers the hardest question about shipping AI into a customer's cloud: when you swap a model per tenant, how do you know you did not make it worse.
See the rest of the work →Clean-room reimplementation from public knowledge. Fixture articles are invented, not any real newsroom's copy. The customer deployment is described without identifying the customer.