CP Tutor: office hours forever

A grounded tutor for corporate training alumni. It answers only from my course materials, in my voice, with citations, so the training keeps working after the room empties.

Status: live
Stack: Python, FastAPI, Cloud Run, Gemini File Search
Links: tutor.corporateprofessor.com

The problem

Corporate AI training has a half-life measured in days. People leave the session motivated, hit the first snag on Tuesday, and stop. The trainer is gone. The slides do not answer questions.

The constraint

A general chatbot does not fix this. It answers from the whole internet in nobody’s voice, and it will confidently contradict what was taught in the room. The tutor had to answer from the course and only the course, say where the answer came from, and admit when the material does not cover the question.

What I built

A retrieval-grounded tutor at tutor.corporateprofessor.com. FastAPI on Cloud Run, with my training materials indexed in Gemini File Search. Every answer is generated from retrieved passages and cites the source document. Questions the materials do not cover get an honest “this is not in the course” rather than an improvised answer.

Access is gated by a cohort code, so it is a service for people who were in the room, not a public chatbot.

It runs in its own Google Cloud project, separate from AI Greg. Keys control access, not scope: any key in a project can read every file store in it, so a shared project would let a compromised tutor key read the personal corpus. Separate projects also mean the budget alarm measures one service at a time.

What it cost or taught

The tutor changed the offer. Training used to end when I left the building. Now the engagement ships with something that keeps answering, in the same voice, from the same material, on the Tuesday when the first snag shows up.