Talks

Available for technical talks in English and Spanish. Based in Madrid, remote-friendly. I speak about building production software with AI agents, evaluating LLMs in production, and making products discoverable to AI agents — every claim anchored to public, verifiable artifacts.

BIO

SHORT

Pablo Albaladejo is an AI Engineer — production LLM systems, agents & infrastructure. At Aircall his systems process 1M+ call transcriptions a day. He created Kaiord, an open-source, local-first training platform written entirely by AI agents — 9 packages on npm, listed in the official MCP registry. With 17+ years in SaaS, he writes and speaks about agentic development, LLM evals, and observability.

LONG

Pablo Albaladejo is an AI Engineer — production LLM systems, agents & infrastructure. At Aircall in Madrid his systems process more than a million call transcriptions a day. Across 17+ years in SaaS he has shipped production backends on AWS and TypeScript, and now focuses on building software with AI agents end to end. He is the creator of Kaiord, an open-source, local-first training platform where every line of code was written by AI agents. Kaiord ships as 9 packages on npm — including an SDK, a CLI, and an MCP server — and is listed in the official Model Context Protocol registry; its evals harness runs 22 benchmarks behind a 90% quality gate in CI. Pablo writes and speaks about agentic development, evaluating LLMs in production, and observability for AI pipelines. Every claim in his talks is anchored to public, verifiable artifacts — code, PRs, and docs. He speaks in English and Spanish and is remote-friendly.

ABSTRACTS

01 45 min · talk

Specs are the new source code

Everyone can get an AI agent to write code. Far fewer can get agents to ship software that survives production. This talk lays out the operating system I use to do exactly that: specs as the source of truth agents work against, a zero-tolerance CI that treats every warning as an error, mechanical guards that catch the mistakes agents reliably make, and human gates where judgment actually matters. Drawing on Kaiord — an open-source platform written entirely by agents and released as 9 npm packages — I show how these pieces fit together, where they fail, and how the human role shifts from writing code to designing the system that lets agents write it safely.

02 30 min · talk

Evals as CI: making stochastic systems boring

"It worked when I tried it" is not a quality bar for LLM features. This talk turns evaluation into engineering. Using Kaiord's open-source evals harness as a running case study, I walk through 22 domain benchmarks, the assertions that grade each response, and the 90% pass gate that blocks a merge in CI when quality regresses. We cover how to choose a threshold you can defend, how to write assertions that catch real failures without flaking, and how to fold evals into the same pipeline as your tests. You leave with a concrete, copyable pattern for treating model quality as a build artifact — measurable, versioned, and enforced — instead of a vibe.

03 25 min · talk

GEO: making your product discoverable to AI agents

Search is no longer the only front door — increasingly, AI agents decide what to recommend. Generative Engine Optimization (GEO) is how you make your product legible to them. This talk covers the concrete surface area: llms.txt as a machine-readable map of your site, structured data and JSON-LD that state facts unambiguously, and the registries — like the official MCP registry — where agents actually look. I use a single-day GEO program I ran on Kaiord as the worked example: what shipped, what moved, and what didn't. You leave knowing which signals are worth the effort, which are cargo-cult, and how to instrument discoverability so you can tell the difference with data.

PROOF

Every claim above is verifiable on these properties.

Invite me to speak

Organizing a meetup, conference, or internal session? A line about your audience and format is enough to start.

Invite me →