This portfolio · AI product · LLM grounding
An AI guide that refuses to make things up
Product owner and builder · 2026 · Independent
The problem
Recruiters skim. A few want to ask a direct question instead — what did he own at Nexorea, where's the resume. A chatbot can answer that in seconds, but an AI that invents a single number in a hiring context is worse than no chatbot at all.
How it works
Suggested questions are answered from pre-written text in the browser, with no model call. Typed questions go to a server route that sends the model only approved portfolio content, and asks for a structured answer that includes a 'supported' flag. Anything unsupported gets an honest fallback and a link to the right page.
- No retrieval or vector database: the whole knowledge base fits in one prompt.
- Placeholder content is excluded from what the model sees.
- Input length cap, per-visitor and daily limits, and a timeout protect cost and uptime.
- No conversations are stored.
Trade-offs
A free model tier keeps cost near zero, but free-tier inputs may be used by the provider to improve its products, and its limits can change. The site states that and stores nothing itself.
Key decisions
No retrieval pipeline
The approved content is a few thousand words.
Chose: Send all approved content every time.
Why: It is the simplest thing that works at this size, and removes a whole class of retrieval errors.