Dan the IT Man
LOCAL AI / A PERSONAL SITE ASSISTANT
A little personality in a green screen. Underneath, a local model helps find the right public fact, and reviewed notes supply the words you read.
Select an allowed fact.
{ "choice": "lab_purpose" }Dan uses his homelab to explore Linux, self hosting and hardware while running services of his own.
Let the model select from approved information. Keep the final wording under editorial control.
01 FROM QUESTION TO ANSWER
Dan-O runs through a Python service and a local Ollama model in the homelab. Familiar phrasing can take a shortcut; other questions ask the model to choose from a limited menu.
The API accepts a bounded question and recent topic/fact identifiers. It checks the request’s shape and limits before matching.
Current question + bounded contextKnown topics and followups use deterministic rules. Remaining questions go to the local model with approved choices.
The service rejects unknown choices, extra fields and model prose. It compiles the result into approved identifiers and a fixed response mode.
Fact IDs, never model-written HTMLThe browser validates the plan again, resolves the catalog text and adds only predefined project links and followup questions.
Public catalog → visible answerThe browser can still answer supported topic names and simple followups from its local catalog. Flexible question matching needs the service.
02 WHAT THAT LOOKS LIKE
Open an example to see the approved response and the decision behind it. These are catalog excerpts, not a live conversation.
Dan uses his homelab to explore Linux, self hosting and hardware while running services of his own.
An exact topic alias resolves to approved facts without model inference. An authored followup can keep the conversation going.
I only have Dan's public notes. I can't provide private information, credentials or internal client details.
The reply stays within the public catalog. The model has no tool for retrieving credentials or inspecting the private inbox.
I'm having trouble connecting to the local model. You can still ask for a topic by name, such as career, homelab or projects.
Supported local replies remain useful. The application bounds inference time and concurrency instead of letting model work accumulate indefinitely.
03 TWO DIFFERENT JOBS
Asking about Dan and sending Dan a support request use separate flows. A model choice cannot submit a ticket or grant consent.
Type an explicit support command to open the guided intake.
Provide the details, review them and explicitly consent before submitting.
The request goes to Dan’s private SQLite inbox. Support fields stay out of model prompts.
Unsubmitted support details stay on the current page and are excluded from the saved public chat snapshot. “New conversation” clears the public session and cancels pending browser work.
04 TESTING THE DESIGN
A reply can use permitted content and still miss the question. Those are different things to measure.
KNOWLEDGE RELEASE · OCTOBER 18, 2026An initial unseen 30-turn comparison improved from 20 to 25 intent matches. A later untouched 12-turn set matched only 7. All replies remained inside the catalog, but some selected the wrong permitted response.
After those misses were corrected, the disclosed set passed 12 out of 12. That became regression coverage; it was no longer an independent accuracy test.
Browser checks covered desktop and narrow layouts, typed responses, fixed terminal size, approved project links, same-tab continuity, reset, rejected provider prose and cancelling support without creating a ticket.
The private inbox was backed up and restored in isolation before the knowledge release. These dated results describe that release; they do not guarantee perfect understanding of future questions.
05 THE TRADEOFF
The catalog makes answers reviewable and keeps the assistant focused on this site. It also means new project details need an editorial update, and a poorly matched question can return an irrelevant approved fact.
Dan-O does not browse the web, run tools or provide general technical troubleshooting. Its job is to introduce my work, help visitors explore it and offer a separate way to request support.
The CRT interface gives it a personality. The validation, fallback behavior and recovery work make it a system I can inspect and improve.