AI Systems
RespondLine
A support desk where the answers come from the customer’s own documents — tickets, live chat and a knowledge base, with retrieval scoped to each workspace.
Built end to end
Built with
- Laravel
- Vue.js
- FastAPI
- PostgreSQL
- pgvector
The hard part
Retrieval that cannot cross a workspace boundary. One embedding store serves every tenant, and similarity search does not know what a tenant is unless the schema makes it impossible to ignore.
What it costs to be wrong
An answer assembled from another company’s internal documentation. It reads as confident, it arrives in a customer-facing chat, and it cannot be recalled.
What was built
- Per-workspace retrieval over pgvector embeddings
- Tickets, live chat and knowledge base under one inbox
- A separate retrieval service, so the model never runs in the request path
- Workspace connectors for calendars, databases and external APIs
Writing
From this system.
-
Nobody has read your AI-generated codebase
It runs, it demos, and no human has traced a request through it end to end. Here is what that costs, and the four things worth checking first.
-
Engineer the boundary, generate the implementation
The useful split is not humans versus agents. It is that the decisions live upstream and the code lives downstream, and only one of those can be handed over.
-
Retrieval that cannot cross a tenant boundary
A similarity search has no concept of a customer. Where you put the filter decides whether that is a design detail or a disclosure, and both failures look like a ranking bug.
More work
Other systems.
Fintech
Taxwift
An income tax filing platform that reconstructs a year of trading from broker statements and files the return against it.
Fintech
Derivative Trading Platform
A real-time derivatives trading system running many managed accounts off one broker connection, with automated execution and risk tooling.
Building something in this area?
Existing ground is faster ground. If your problem sits near work we have already done, we can be useful quickly.