Presenting Your AI Portfolio

"Tell me about something you've built with AI." For a .NET + AI role, this question decides the interview — and a real, running side project beats any certificate. Suppose your evidence is an AI learning platform you actually built: a two-course site with a RAG tutor that answers from lesson content with citations, AI-generated quizzes, local and cloud model routing, and SQL Server vector search. This lesson is how to turn that from "a thing I made" into interview evidence.

The two-minute tour

Rehearse a tour with this shape: problem → architecture → the hard decision → the result. "I built a learning platform with an AI tutor. When a student asks a question, the app embeds it, runs a vector search over the lesson content in SQL Server, and sends the top chunks to the model with instructions to answer only from those sources and cite them. If retrieval comes back weak, it refuses instead of guessing. It routes between a local model for cheap requests and a cloud model where quality matters. It's live, with real users." Under two minutes, no jargon dump, ends on live with users — the phrase that separates you from tutorial-followers.

The decisions worth highlighting

Interviewers reward decisions with reasons, not features. Lead with these:

• Fail-closed refusal — the tutor refuses when retrieval confidence is low, because in an education (or government) context a confident wrong answer is worse than "I don't know." This maps directly onto public-sector risk tolerance; say so.
• Citations by construction — answers must quote their sources, which makes hallucination visible and auditable. The word auditable lands hard in an agency interview.
• Local/cloud routing — cost, latency, and data-sensitivity trade-offs per request type; exactly the conversation agencies have about which data may leave their tenant.
• Vector search in SQL Server — you chose boring infrastructure the team already operates over a new vector database. Justifying not adding a component is a senior signal.
• Evals — you tested retrieval quality and answer behavior against a question set rather than eyeballing it. Even a modest eval habit puts you ahead of most candidates.

Inviting deep-dives you can win

End the tour with a door: "happy to go deeper on the retrieval pipeline, the refusal logic, or the eval setup." You've just steered the follow-up onto your strongest ground. Before the interview, rehearse the second level of each: chunking choices, how you picked the refusal threshold, an eval case that caught a real regression, one thing that's still weak and how you'd fix it. Volunteering a known weakness — "quiz generation still needs human review; here's why" — builds more trust than pretending it's finished.

Red flags

• Feature-listing without a single "because."
• Claiming the AI "never hallucinates" — you'll be dismantled by the first probing question.
• Being unable to answer "what breaks if the model is down?" about your own system.

Practice prompts

1. Deliver your two-minute tour aloud, timed. Cut it until it fits with room to breathe.
2. "Why didn't you use a dedicated vector database?" Answer in under 45 seconds, trade-offs included.
3. "How do you know your tutor's answers are grounded?" Describe your eval approach and one concrete failure it caught.