Zamana

Zamana is a personal shopping assistant backed by a product recommendation engine. While I was there, it was mainly something you’d text on iMessage, like a friend who’s an expert in every consumer product category.

I spent six months there building the harness for the agent behind it.

What I Worked On

The harness is everything around the model that turns it into a product: how a conversation is turned into context, which tools the agent can call to search and compare products, and how its recommendations get back to the person on the other end of the chat. I built it from scratch, then spent the rest of my time maintaining and optimizing it.

Most of that work came down to making the agent fast and reliable:

• Context engineering: deciding what the model sees on every turn (conversation history, user preferences, product data) so it has what it needs without drowning in it
• Speed: in a chat on iMessage, a slow reply feels broken, so latency mattered at every step of the loop
• Reliability: an agent that recommends products has to be right consistently, not just in demos, so a lot of the work was finding and fixing the ways it failed