01 / The Work
Client Roadmap And AI Docs Platform
RBAC Kanban Roadmap · Embeds In Any Host App · Docs Written On Every Merge · Delivered 2 Weeks Early
Built At Antz →The Problem
Bug reports and feature ideas kept dying somewhere between the person who had them and the team who could act on them. Docs had the opposite problem: nobody wrote them, so they were always behind the code. The client wanted both fixed, across several apps, without asking anyone to learn a new tool.
The Decisions
One Roadmap, Many Host Apps
Instead of a roadmap per app, it's one central service that an admin drops into each host app as a page. Same data, same rules, and a user who touches three apps sees all three in one place.
RBAC Decides What You See
Access is scoped per app. You see the asks and progress for apps you have access to, and nothing else. That made it safe to share one board across teams that shouldn't see each other's work.
Docs Follow The Code, Not The Calendar
Docs are rebuilt from the merge itself, triggered by a service hook on main. If the code changed, the docs changed. No one has to remember.
One Database For Everything
Data lives in Lakebase Postgres on Azure Databricks, part of a central data platform our team built for the client. Their SQL Server databases were retired and the schemas moved over, so this app reads and writes the same managed data as everything else.
Never Trust The First Draft
Model output goes through a review loop before anyone sees it, with three tries and an alert if it can't pass. It's validated like any other untrusted input.
How It Works
- 01Spot a bug or want a feature? Raise it on the roadmap board inside the app you're already using. It lands on a Kanban board where it gets tracked, not forgotten in a chat thread.
- 02The roadmap is built once, centrally. An admin onboards it as a page into any host app, so one person can follow asks and progress across every app they have access to, and see nothing they don't.
- 03Merge a PR to main in any project repo and an Azure DevOps service hook kicks off the docs pipeline. An OpenAI model hosted on Azure AI Foundry writes the user guide and the technical docs.
- 04The draft goes through a review loop with up to three attempts. If it still fails, or anything else breaks, an alert goes out. The roadmap side has its own alerting too.
- 05Readers get a sanitised docs view with an index, citations, section reference links and navigation. Nobody writes any of it by hand.
- 06Under the hood it's all Azure: Next.js on the front, a Hono API, Lakebase Postgres on Azure Databricks for data, Blob Storage for attachments, Azure Communication Services for email, Entra ID for SSO and OAuth, and Azure DevOps Pipelines for CI/CD.
High-Level Design
The Outcome
Shipped to the client two weeks ahead of schedule, even with blockers that sat on other teams. The generated docs score around 85% on cohesiveness and are far easier for end users to follow. Next up: a Playwright integration that captures screenshots and places them in the right spot in the docs.
~85% cohesiveness on generated docs · 3-attempt review loop · alerting on both halves
Two Weeks Early. Blockers Included.