Artificial Intelligence Engineer
Pubblicato il 12-09-2026 - TextYess in Roma
We'rethe AI-first CRM that turns conversations into revenue. On-site, WhatsApp, email, voice: our AI Agents sell, support, and retain customers for the best B2C brands out there, autonomously.
We believe the future of commerce is warmer, more human than ever - just at a scale no human team could pull off alone.
We've raised €3M+ and work with brands like Ducati, Pittarosso, MC2 Saint Barth, and Doppelgänger. We're a lean, fast, Italian-born scale-up now expanding our story internationally and this role is a big part of writing it.
The role
Join our AI team to build the agentic runtime for proactive agents: context building, tool/retrieval orchestration, scoring, and guardrails. You’ll own production components from day one and work directly on the “signal vs. noise” call.
What you'll actually be doing
Grounding. The agent sometimes says a product is unavailable when it's in stock. Fixing that is a retrieval problem, a prompt problem, and an eval problem at the same time.
Exact match vs. semantic. Semantic search is good at "warm jacket for winter" and bad at a SKU, a size,
or a specific variant. We need both, in one ranking.
Multilingual quality. Our customers write in Italian, English, and more. An agent that answers in the wrong language is worse than one that says nothing.
Latency and cost. A customer is waiting on their phone. You'll profile requests, test prompt and model changes, and ship improvements you can point at on a graph.
Evals that catch problems before customers do. Every conversation has a Langfuse trace, and we have a golden-dataset harness. Making it actually predict production quality is open work.
Deciding what's worth saying. Ranking, scoring, and suppressing proactive insights so the agent interrupts a customer only when it has earned it.
What you bring
2+ years shipping software professionally , including at least one LLM feature real users touched — a chatbot, a RAG system, an agent. Not a weekend project.
You know how RAG actually behaves :