What's happening here?
This page explains the experiment running at aigreg.bot. It is written for students, colleagues, and the curious. It updates automatically: the page is part of the code, so every time the system changes, this explanation ships with it.
The short version
Greg Dyche teaches management information systems, analytics, and AI at Creighton University. Instead of only talking about AI agents in class, he built one. aigreg is Greg's AI front door: it answers questions about Greg, his courses, and his work, using only things Greg has actually written. When it cannot answer, it does not guess. It hands the question to the real Greg and tells you he is on it.
You are welcome to watch this being built, poke at it, and learn from what breaks. Things break regularly. That is the point.
One brain, many doors
There is one knowledge base and several ways to reach it.
- The web page. The chat box at aigreg.bot. No account, no setup.
- Email. Write to email@aigreg.bot and a reply comes back in a minute or two. Replying continues the conversation.
- Your AI agent. If you use an AI assistant that speaks MCP (the Model Context Protocol), point it at https://mcp.aigreg.bot/mcp and your agent can ask Greg's agent questions directly. Agent-to-agent communication is not science fiction; it is a config setting.
Every door leads to the same brain, called the corpus: a set of documents distilled from about a decade of Greg's own writing, his articles, his podcast scripts, his course materials. The AI is only allowed to answer from those documents. This matters more than it sounds. It is the difference between an assistant that represents someone and a chatbot that improvises.
How answering actually works
When a question arrives, the system searches the corpus and drafts an answer from what it finds, along with a confidence signal. Confident, grounded answers go out. Everything else follows one rule, and it is the most important rule in the system: never bluff.
A question the corpus cannot answer becomes a ticket. The asker is told, honestly, that this one is beyond the AI and has been passed to the human. Greg reads every ticket himself and typically answers within a day. His answers then become candidates for the corpus, so the system learns, but only through Greg's own editorial hand. Nothing enters the brain automatically.
Trust is earned in tiers
Anonymous visitors get answers from public material. Verify an email address (the system sends a confirmation link) and deeper material opens up. This is the same idea as any relationship: you learn more about someone as you introduce yourself.
The guardrails
Two recent lessons produced two guardrails, and both are honest teaching material.
The student privacy rule. aigreg can discuss anything class-level: the syllabus, course tools, policies, how a course runs. It will not discuss anything individual-level: a specific student's grades, absences, extensions, or circumstances. Those questions get an immediate redirect to official channels, and deliberately leave no trace in the system. The strongest guardrail is not a clever prompt; it is that the AI has no access to the gradebook at all. You cannot leak what you cannot see.
The broken promise rule. Early on, someone emailed asking for a bio and a photo. The AI answered the bio beautifully, then wrote "I will have Greg send over a photo." No code backed that sentence. No task existed. The promise lived only in the text. The fix was twofold: the AI is now forbidden from promising actions, and the system detects requests that need a human (send a file, schedule a meeting) and opens a real ticket even when the informational part was answered. Lesson: an AI saying it will do something is not the same as a system doing it. Sit with that one. It generalizes.
The course corpus
BIA 253 (Management Information Systems) students can ask aigreg about the course: policies, tools, how things work. Two design rules keep this sane. First, the corpus holds ideas, not numbers: grading scales, dates, and rates live in the syllabus on Blueline, the single source of truth, and aigreg points there for specifics, so nothing here can silently go stale. Second, the privacy rule above applies always. Ask about your own grade and you will be politely sent to Professor Dyche himself.
The stack, for the technically curious
Python. A FastMCP server on Google Cloud Run. Retrieval via Gemini File Search over the corpus store. Tickets and conversation state in Firestore. Email in and out through a dedicated Gmail account the machine owns. The web page you are reading is one Python string, no framework, no build step. The whole system deploys with one command and is small enough that one person can hold all of it in their head, which is a feature, not a limitation.
Lessons so far
- Never bluff. Low confidence routed to a human beats a fluent wrong answer.
- An AI's promise is not a system's action. Detect requests, create tickets.
- Do not duplicate facts. One source of truth, everything else points to it.
- Fail closed. When a security filter cannot be applied, return nothing.
- Guardrails made of capability beat guardrails made of prompts.
- The conversation is the point. Escalation is not failure; it is the product working.
What's next
Texting and WhatsApp as new doors. Calendar integration so scheduling becomes real instead of escalated. More course corpora as more classes come online. This list will change, and when it does, this page will already know.
Try it
Ask a question in the chat at aigreg.bot. Email email@aigreg.bot. Point your AI agent at https://mcp.aigreg.bot/mcp. BIA 253 students: ask it about the late work policy, or whether you have to use Frame. And if you find a way to break it, tell Greg. He reads everything, and he will probably thank you in class.