I run support teams. I build the AI they use.
These are working demos, not screenshots. They run in your browser. Nothing you type leaves this page.
Loading example…
-
Prompt builder
Build a prompt from parts. It gets checked live against the mistakes that make support prompts fail.
Look for: missing output format, no way to say "I don't know", untagged customer text.
-
Tokens and cost
See how a model splits text into tokens, then what that costs per month across 12 current models.
Look for: wallet addresses and older tokenizers on Japanese or Tamil. Same meaning, several times the bill.
-
Embeddings
A small model runs on your device and turns sentences into vectors. Compare them, map them, search them.
Look for: how close "can withdraw" and "cannot withdraw" score. That is why RAG needs more than vectors.
-
Support RAG
Answering from a help centre with citations, a retrieval inspector, and a switch that breaks it on purpose.
Next.
-
Evals and guardrails
Golden questions, pass rates before and after a prompt change, prompt injection hidden in a ticket.
After that.
Why it is built this way
No API keys, no backend, no third-party scripts or fonts. The models and libraries are files on this server. Your browser downloads them once and runs them.
That keeps the cost at zero, and it keeps customer-style text off other people's servers. In regulated support work that second point matters more than the first.
Who built it
All data on this site is made up. No employer code or data is used.