Satya Nadella Put This AI Tutor on Stage. I Rebuilt It in a Weekend.

In 2017, a lecturer at UNSW built an AI tutor so good that Satya Nadella put it on stage.

David Kellermann built QBot. A chatbot inside Microsoft Teams that answered student questions, routed the hard ones to human tutors, and tracked response quality across entire course cohorts.

Student satisfaction in his engineering courses went from 74% to 99%.

Microsoft flew Nadella to Sydney to keynote it at Innovate 2018. They wrote it up as a flagship education case study. It was real, it worked, and it mattered.

The technology underneath has mostly been retired since then. Bot Framework V3 is legacy. QnA Maker shut down in 2025. The PowerShell orchestration layer was always a prototype that survived longer than anyone expected.

David and I are now rebuilding it together with the 2026 stack.

What I Did Last Weekend

I rebuilt it. One developer, 36 hours, $200 in free Azure credits, using the 2026 Microsoft AI stack.

35 Azure Functions. A Cosmos DB knowledge graph with 80+ vertices mapping students, courses, topics, and learning outcomes. 494 documents indexed in Azure AI Search. Foundry Agent Service handling the orchestration layer. 20 demo students across 6 courses. Three Teams tabs for course admin, gradebook, and teacher insights.

Running cost: under $5 per day on infrastructure that most universities already pay for through M365 A5 Education licensing. Total spend so far: less than what most university IT departments spend on a single planning meeting.

I called it CohortAI. It does everything QBot did, plus a few things that weren't possible in 2017: automated early warning detection when students start falling behind, curriculum analytics showing which topics are causing the most difficulty across a cohort, and cross-course intelligence that connects patterns between departments.

Upload a PDF course outline. The system parses it, builds the knowledge graph, provisions the Teams channels, and configures the AI tutor with course-specific context. Any course. Any university. Five minutes.

The Honest Numbers

I logged every mistake during the build. There were 20 of them.

They cost roughly 10 hours of the 36-hour build. That means 28% of my time was spent on problems I created for myself or walked into because the documentation was wrong, outdated, or buried in a GitHub issue comment from 2025.

A toggle in Power Apps that looks identical in two modes but silently changes how formulas are interpreted. A configurable timeout setting that Azure Functions Consumption plan simply ignores. A Cosmos DB API version that exists nowhere in official documentation but works perfectly if you know the string.

These are the kinds of problems that turn a weekend project into a three-week project for someone encountering them the first time.

Why I'm Publishing Everything

I wrote a grounding essay this week that covers the full story. David Kellermann's original work. What the 2026 stack actually looks like in practice. The architecture decisions. The real costs. Where I'd do things differently on a second pass.

Starting this week, I'm also publishing the complete build guide as a series. Each piece covers one stage of the build, from infrastructure provisioning through to the knowledge graph, the AI Search pipeline, the Foundry Agent Service configuration, and the Teams app deployment.

Every Azure Function documented. The knowledge graph schema explained. All 20 fuckups catalogued with what went wrong and how to avoid them.

I'm keeping nothing back. The architecture diagrams, the deployment scripts, the mistakes, the workarounds, the cost breakdowns. All of it, published, free to read.

Who This Is For

If you work in higher education IT and you've been quoting six-figure budgets for AI tutoring pilots, I'd like you to see what $5/day looks like.

If you're an educational technology vendor charging per-seat licensing for capabilities that now ship inside M365, your customers are going to start asking questions.

If you're a course coordinator who was told “AI tutoring is too expensive and too complex for your faculty,” you deserve to see the architecture diagram before accepting that answer.

If you're a CTO or CIO at a university and you want someone to build this for your institution rather than figuring it out yourself, that's what my consultancy does.

The Guide Series

The grounding essay publishes first. It covers the story, the stack, and the reasoning behind the architecture choices.

Then the build guides drop one at a time across the week:

• Azure infrastructure and provisioning
• Cosmos DB knowledge graph design
• AI Search pipeline and document ingestion
• Foundry Agent Service and orchestration
• Teams app integration and deployment
• The 20 fuckups (what went wrong, what I'd skip next time)

Each guide is written so that a competent developer with Azure experience can follow along and build their own version.

What I'm Not Publishing

My consulting rate.

Everything else is in the guides. If you want to build it yourself, you'll have everything you need. If you want someone to build it for you, scope it for your institution, and handle the integration with your existing systems, that's a conversation worth having.

I work with universities and education providers on AI integration through Frontier Operations. If any of this is relevant to what you're planning, reach out.