Building AI Agents
Today, anyone can build an AI agent demo that looks great. The hard part comes after: making the product actually useful, keeping it reliable, and getting it a bit better as your user base grows.
Demos run on a handful of examples someone picked by hand. Production is messier. The same product shows up ten different ways across sources, prices move by the hour, and a feed can go stale for days before anyone notices. Your users don’t see any of that. They see a wrong answer, and they stop trusting the agent pretty fast.
We don’t think the teams that get ahead are winning on prompts. They have a loop: find where the agent gets it wrong, figure out if it’s a data problem or an agent problem, fix it, and do that at scale. There’s AI on both sides of that loop, too. LLMs clean up and enrich the raw data going in, and LLMs grade the agent’s answers coming out. In this event we're gathering three companies working on this:
At Joko we’re building Juno, a shopping assistant that sits on top of a catalog of hundreds of millions of products. If it gets a price or a stock level wrong, or simply doesn’t have the product, nothing else it does matters much.
Perk does AI for business travel and event booking. Tell someone a venue is free when it isn’t, or quote the wrong price, and you’ve broken their plan.
ListingLeads is changing how real estate agents do email marketing. An agent won’t hit send on an email with the wrong listing details or old market numbers, so the data has to be right.
Program
18:15 — Doors open
18:30 — Panel with:
Mike San Román — Building ListingLeads, prev. ex-Director of Engineering at Buffer
Dan Neciu — Engineering Manager & Staff Engineer @ Perk
Alexandre Djerbetian — Software Engineer @ Joko
19:00 — Q&A
19:15 — Drinks, snacks & networking 🍻
This is our first engineering event in Barcelona, organized by the Joko team here. If you’d rather talk about evals, data pipelines and the unglamorous work after launch than about launch day, come join us.