Trustworthy AI
The AI assistant we run our own business on
Our own system, built by Raghav for our own use.
A permission gate that has checked 1,075 requests from an AI assistant and refused 38, with every action check recorded in a hash-linked log.
- Requests checked by the permission gate
- 1,075
- Refused
- 38
- Permission grants, every one revoked when its job was done
- 193
The situation
Before letting AI take actions for clients, we wanted to run it on our own work first: operating a computer, changing its own memory, and answering from our own documents.
What was hard
An assistant that can act can also act wrongly. The goal was to make it useful without giving it standing access to anything it doesn’t need.
What Raghav did
- Built a broker between the AI and its tools. It refuses unknown tools, allows writes only when authorised for the current request, keeps credentials out of the model’s view, and shuts off after three refusals in a row.
- Recorded every action check in a hash-linked log, so entries can’t be quietly changed after the fact.
- Capped every permission grant at a limited tier and revoked each one when its job was done.
- Built a knowledge base of 875 documents across 130 topics that the assistant answers from, citing its sources.
- Tested a new search method before shipping it, and held it back when it wasn’t ready.
What changed
The gate has checked 1,075 requests: 847 actions (18 refused), 178 memory changes (5 refused) and 50 work requests (15 refused). Most action checks are automatic ones made during computer sessions a person started, not separate human approvals.
All 193 permission grants have been revoked, and the emergency stop has never been needed. The new search method scored better in testing (mean reciprocal rank, a measure of how high the right answer ranks, rose from 0.41 to 0.56), but it wasn’t shipped because it wasn’t yet running reliably.
How this was measured. Counts come from the permission gate’s own logs. Action checks cover September and October 2026.
Bring us in when you want AI that takes real actions in your business, with limits you can see and a record of everything it did.