I speak on making AI work in production: evals, observability, cost-efficient scale, and the agentic development practice behind it. Every talk is concrete and shows the proof. I give talks in English and Spanish. I am based in Madrid and can also speak remotely.
SHORT
Pablo Albaladejo is a Senior AI Engineer at Aircall, where he makes AI work in production. There he keeps it predictable in quality and cost, on a platform that handles more than a million calls a day. He is the founder of Deepgent, a multi-tenant AI chatbot SaaS. He also created Kaiord, an open-source, local-first training platform that agents wrote end to end.
LONG
Pablo is a Senior AI Engineer (AWS, TypeScript) with 17+ years in software engineering. At Aircall he builds the AI behind call summarization: real-time transcription and LLM features on serverless microservices. The platform handles over a million calls a day. His job is to make that AI predictable in quality and in cost: evals as CI gates, traces on every call, and cost-efficient scale. He is also the founder of Deepgent, a multi-tenant AI chatbot SaaS that is live in production. He builds his products end to end with agents, under spec-driven development and a zero-tolerance CI. One of them is Kaiord, an open-source, local-first training platform with a TypeScript SDK, a CLI, and an MCP server.
He speaks to engineering audiences on treating evals as CI, tracing stochastic systems, and briefing agents with specs. He is AWS certified at Professional level. He also holds two certifications: Claude Certified Associate - Foundations and GitHub Certified: Agentic AI Developer. He is based in Madrid, Spain.
01 25 or 45 min · talk
This talk covers why offline benchmarks lie, how to resample golden sets from production, and how to gate deploys on evals. With that gate, model changes ship weekly instead of quarterly. It draws on running LLM features on a platform that handles 1M+ calls a day.
02 30 min · talk
This talk shows how to brief AI agents so they ship production code: the spec format, the review loop, and the CI that catches what the agent misses. It also covers what changes when the bottleneck moves from writing code to specifying it.
03 25 min · talk or workshop
Aggregate metrics hide the failure, and the trace names it. This talk shows how to instrument retrieval, tool calls, and token flow, so a regression bisects to a diff in minutes instead of a long manual search.
04 25 min · talk
Search is no longer the only way people find a product. More and more, AI agents decide what to recommend. Generative Engine Optimization (GEO) is how you make your product readable to them. This talk covers the concrete surfaces. llms.txt gives agents a machine-readable map of your site. Structured data and JSON-LD state facts clearly. Registries like the official MCP registry are where agents look. The worked example is a one-day GEO program I ran on Kaiord: what shipped, what moved, and what did not. You leave knowing which signals are worth the effort, which ones only look useful, and how to measure discoverability so you can tell the difference with data.
You can check every claim above on these links.
I can speak at conferences, internal engineering days and podcasts. I adapt the talk to your audience and bring the proof.