Jeremy Neely Text Jeremy

AI & Agents · Production

Secure AI Hatz Tips

Fact-checked how-to answers inside the AI chat customers already use.

Built for: Business customers of an AI onboarding program, about 600 end users

Tips as customers see them: short steps, why it helps, and the source.
Tips as customers see them: short steps, why it helps, and the source.

The problem

Customers learning Hatz AI kept asking the same how-to questions, and the answers were scattered across decks, help articles, release notes and dozens of vendor videos. Hand-written tips do not scale, and AI-written tips are often wrong: about 38% of the first batch pulled from videos contained a claim the video never made.

What I built

I built a tips service that customers call as an MCP tool from inside their own Hatz chats, and get short answers that link to the exact article or video moment. Behind it is a .NET platform on a Linux server that turns vendor videos and docs into fact-checked, de-duplicated draft tips for review, plus a supervisor that watches and heals the system.

How it works

  1. Pull new vendor videos and help pages on a schedule.
  2. Have AI draft short tips with timestamped links.
  3. Fact-check every tip against its source and drop duplicates.
  4. A person approves tips in a web console.
  5. Customers ask in their Hatz chat and get cited answers.

Built with

  • C# / .NET 10
  • ASP.NET Core
  • MariaDB
  • Claude Code CLI
  • Hatz AI
  • Ubuntu Linux

More of Jeremy's work