① The problem.
A portfolio gets a thirty-second skim — index, bio, a project or two, gone. I wanted a way for the hiring manager who's actually interested to ask the real questions — "what's his IT/OT experience?", "is he a fit for a Solutions Architect role?" — and get a straight, grounded answer instead of more prose to scroll past. The catch is that a bot which invents a project, or claims a skill I don't have, is worse than no bot at all. The whole thing only earns its place if it's honest, so honesty is the part I actually engineered.
② Approach.
The corpus — my projects, bio, résumé, the /now page, plus one hand-written narrative — is about eleven thousand tokens. That fits in a modern context window whole, so there's no vector database and no retrieval step to get subtly wrong: the site assembles the entire prompt at build time — system rules, the full knowledge pack, the conversation — and hands it to a model whose only job is to generate the next answer. Everything that defines the bot's behavior lives in my repository, version-controlled, not baked into weights or scattered across an external index.
③ Keeping it honest.
Three bounds, in order of how much they matter:
- It answers only from the provided content, and says "I don't know — here's the contact page" when the answer isn't there. No outside knowledge, no guessing about salary or availability.
- It declines anything off-topic. It isn't a general chatbot; it'll politely refuse to write your poem and point you back to what it can actually cover.
- It never names my current employer — and that one isn't enforced by asking the model nicely. A build step scans the assembled knowledge pack and fails the build if the name ever leaks in, so the model literally can't reveal what it was never handed. I spent an afternoon trying to break it — direct questions, "ignore your instructions," a forged system message claiming the rules were switched off — and it held every time, on every model I ran.
④ The backend call.
I built this to run on my own hardware. Switchboard, my GPU broker, sits in front of a 16 GB Tesla P100 in a closet and was meant to host exactly this. Then I measured. Prefilling an eleven-thousand-token prompt on a Pascal-era card takes about ten seconds warm and the better part of a minute when the model has been evicted — and the small local models that fit either followed the bounds loosely or, in one case, were quietly being fed a truncated prompt because the default context window was a fraction of what the pack needed. A larger local model grounded well once I fixed the context, but still wandered off-topic more than I'd ship.
So I made the call I'd make for a client: the live assistant defaults to Claude Haiku, which answers in under a second with the static pack prompt-cached, and the P100 path stays wired as an option and a fallback. The provider is one environment variable. Choosing the right tool for the constraint is the job; running it on my own GPU just to say I did isn't.
⑤ What it is, and isn't.
This is home-grown — its own small codebase, not the hermes-agent framework that drives the dashboard on this site, and not Cervi, the assistant I built at work. It's the same shape as Cervi — a bounded, retrieval-grounded assistant with a curated knowledge base and system-prompted bounds — rebuilt here in the open so you can poke at it instead of taking my word for it. Which is the whole point: scroll up, open it, and try to catch it in a lie.