Overview
A production assistant that lets people run a programmatic ad platform by talking to it instead of clicking through configuration screens.
Why it exists
Running an ad platform means living in dense settings pages. Something simple, like tweaking a rule or checking how a campaign did, turns into a dozen clicks and a fair amount of tribal knowledge. I wanted that to be a sentence instead.
How it works
It is a stateless Python service that sits between a language model and the platform's REST APIs. The model calls tools, the service runs them, and answers stream back to the user token by token over SSE.
A few things I cared about while building it:
- Parallel tool calls that actually work. The model often asks for several things at once, so the service reassembles those calls as they stream in and feeds the results back over multiple rounds.
- It asks first. Anything that changes data waits for a confirmation, and every change can be undone. That was non-negotiable before I let it near real campaigns.
- Answers you can read. Raw APIs hand back IDs. The service quietly swaps them for names using a cached lookup, so replies read like a person wrote them.
- Analytics when they are there. Reporting tools get added to the toolset per request, and the assistant keeps working normally when they are not configured.
- Answers grounded in the docs. I'm currently adding a RAG pipeline over the platform's large documentation set, so when someone asks how a feature works, the assistant retrieves the relevant pages and answers from them instead of guessing.
Where it landed
Work that used to need someone who knew the UI inside out is now just a conversation, with enough guardrails that people trust it to make real changes.