AI website assistant
2026 · Part of this site
The chat on this site
Answers questions about my services with a Nemotron model hosted by NVIDIA, and a local open-source model as backup.
Context
I wanted a site assistant that answers in seconds, stays within what I have published, and tells me what people are asking.
What I built
- 01Each message and a bounded amount of recent context are stored on Cloudflare and sent to NVIDIA's API
- 02The Nemotron 3.5 Super VL model writes the reply from the offers and prices published on this site
- 03If that service does not respond, the message waits in a secure queue and a Gemma model, run locally through Ollama on my computer, answers it
- 04After each processed turn, a summary reaches me by email
- 05Sessions are anonymous, protected by Turnstile and limited to two hours and twelve turns
- 06Messages are kept for 30 days and deleted earlier on request
Stack
- Nemotron
- NVIDIA API
- Gemma
- Ollama
- Cloudflare Pages Functions
- D1
- Turnstile
- Resend
What it shows
An assistant can answer quickly from a hosted model and keep a private one in reserve. The key stays on the server, and the visitor is told which provider answers.
Tell me what eats your week.
Send a few lines about the task that keeps coming back. I reply within one working day, usually in the evening, Madrid time.