AI regulatory content pipeline with expert review gates

A professional-services firm wanted regulatory commentary without its experts writing first drafts. Every claim needed a recent primary source, and reviews take days.

Solo: architecture, code and prompts, and the deployment on Azure with Terraform.

• Architecture: Durable Functions, with multi-day pauses, fan-out and retries built in; models, embeddings and publishing behind adapters, so a provider can be swapped.
• Frontend: Reviewers approve topics and translations on server-rendered pages. One click opens a draft in Word for the web, and edits save back when they close it. An AI revise option rewrites a draft on request.
• Backend: Parallel drafting, escalation of stalled reviews, duplicate-run refusal and a halt on conflicts of interest.
• Data & AI: Daily ingestion with OCR and a 45-day freshness window; hybrid search on Azure AI Search with a domain-specialised Voyage AI embedding model, chosen for the client's material; versioned prompts.
• Security: Signed, expiring review links, a nonce-based CSP, managed identity, and logs that refuse tokens and draft text.
• Infrastructure & DevOps: Terraform with 55 resource definitions and 18 alert rules.
• Testing: About 1,600 unit tests and 16 verification scripts.

• Structured output is for data, not prose. My first design asked the model for the whole article inside one JSON answer, and drafts came back around 300 words. I traced it to the format, not the prompt, and split writing from structuring into two calls: the median rose to 1,084 words, enough for experts to edit rather than rewrite.
• A prompt rule is not a control. Some company names must never appear, so a deterministic check backs the prompt rule. Tested on the firm's back catalogue, it caught the one real mention with no false blocks.
• Measure before building. Before building spend controls for embeddings, I priced a run: they were 7–10% of cost, OCR about 90%. I fixed the real bug instead.

Experts get cited drafts to review instead of writing them. A production-configuration run passed 28 of 28 end-to-end checks, from conflict screening to per-article approval in both languages.