The product: an AI tutor for the Italian National Tour Guide Exam, built for a licensing exam that requires candidates to speak, unscripted, in front of an evaluation panel. There was no way to rehearse that, so I built the examiner.
The problem, and why it's an SME problem
Strip away the exam-specific detail and it's a problem most SMEs have in some form: a skill that can only be learned live, in front of a real evaluator, where practice is scarce and feedback is slow. Sales pitches rehearsed only in your head. Frontline staff corrected once a quarter, after the mistake. Any process where reading the manual and performing under pressure are different skills, and only the second one gets judged.
The same logic carries over directly to client work. The AI doesn't need to replace the salesperson or the frontline employee, it needs to replace the missing rehearsal: patient, always available, reacting to what someone actually said instead of a script. That's usually a narrower, cheaper, and more defensible use case than "AI does the whole job."
What I actually built
- Voice capture & live transcription, 5–7 minutes of speaking, browser-only, no app.
- An LLM grading engine scored against the real official exam rubric, using the same criteria as the actual panel.
- Follow-up questions generated from the candidate's own answer: the one part of the exam that can't be rehearsed any other way.
- A weak-point map that flags exactly which part of the answer was thin.
- Tiered access (free, standard, full) with a server-enforced free quota, not a time-limited trial.
- Payments, reconciliation and refunds wired directly to entitlement logic, with no manual reconciliation.
product walkthrough
The decisions that matter more than the tech stack
Anyone can call an LLM API. The judgment is in what you build around the call, the part that transfers directly to any AI project I'd take on for a client:
- Adaptive content, not a content library. Follow-ups and the weak-point diagnosis are generated per session from what the user actually said.
- Privacy minimized by default. No audio is ever stored, and transcripts exist only for the duration of the call.
- Built and shipped solo, using Claude as the development partner itself: the same leverage available to any small team directing it well.
Every one of these transfers directly to client work. The tech stack changes, the judgment calls don't.