Michael Mortensen

A RAG demo shows one good answer and stops. I wanted to know how often the right passage is retrieved, when the service should refuse, and whether a change made things better or worse.

A C# ingestion pipeline into Azure AI Search, an ASP.NET Core answer API that cites its sources and refuses when they fall short, a React UI, and an MCP server. An eval harness runs in CI against Azure, and the whole stack is defined in Bicep.

Hybrid retrieval put the right doc in the top 3 for 93% of 40 hand-written questions, against 78% for keyword-only. All 3 unanswerable questions were refused in every run. The question set is small, so a few points either way is noise.

Most starters stop where the hard part begins. Every new business app ends up rebuilding auth, layout, and plumbing for the tenth time.

A production-shaped dashboard app you can clone, rename, and build your real product on: React and TypeScript, an ASP.NET Core API with hand-written SQL through Dapper, and schema changes as plain SQL scripts applied by an explicit migrate command. Every significant decision gets written up in a post.

Live at dashboard.mvmorten.com, deployed to Azure on every merge once CI has checked the build and the schema against a real SQL Server. The API connects with its own login that can read and write data but can't change the schema.

Employees googled assisted-living facilities one at a time and guessed at emails. That approach couldn't get past a few hundred records.

A pipeline that collects records from government sources in all 50 states, enriches each through Google APIs, and validates every email for deliverability, with no human in the loop.

Nearly 50,000 facilities structured, 11,500+ verified emails, and roughly 4 person-years of manual research replaced.

The front end was a set of prototypes built by an outside team who knew the API but not React. It was full of dead code, had no consistent structure and no tests, so every change risked breaking something else.

I removed the dead code and reorganized everything into a coherent codebase, with consistent component and state management patterns the whole team could build on. I also built the front-end unit test suite from scratch. The first three months were hand-written; the last six were driven by Cursor and Claude Code.

Front-end test coverage went from 0% to 98%. The codebase became much clearer to work in, with far fewer regressions.

Hourly staff hand-keyed fund and marketing data into 30 partner sites that had no API. Every typo was a credibility problem, and every departure meant retraining someone on all 30.

An analyst-approved pipeline that logs into each site, publishes through the browser, then verifies every push in a fresh session, with a screenshot of each step.

1,000× output, all 30 sites updated automatically, manual-entry errors eliminated, and 5× revenue growth over two years.