Practical guides on AI engineering, from production RAG systems to voice agents. Real lessons from real projects.
Osvaldo Restrepo — AI engineer, PhD, and full-stack developer based in Panama — shares what it's like building production AI systems from Latin America, why Panama is an emerging tech hub, and what the AI engineering landscape looks like in Central America.
The story behind building an AI system for NICU families. Born from lived experience, grief, and the realization that parents deserve better tools to advocate for their children.
Engineering AI features for people who don't know — and shouldn't have to know — what an LLM is. Five rules earned from MILA, voice agents for small businesses, and watching kids try AI tools at Waldo's Code Lab. Say what it can't do on the first screen, hand off every output, treat latency as a feature, suggest examples that prove the boundary, and never fail with the word 'model.'
The flashy agent demos are also the easiest to misbuild in production. The loop — model calls a tool, gets a result, calls another, repeats — needs strict termination criteria: max iterations, no-progress detection, token and time budgets, an explicit escape hatch. Why letting the agent decide when it's done is the bug that ate your invoice, and the boring patterns that make loops survive contact with real traffic.
In clinical settings, a wrong answer does not have one cost — it has a tone. A confidently wrong answer slips into the chart. An uncertain wrong answer gets double-checked. Why a 95% accurate, always-confident model is more dangerous than an 88% accurate, properly calibrated one — and how I build calibration into the surface of MILA.
Slapping 'AI' on a feature changes user expectations in ways your finance team didn't price in: refund liability, a higher reliability bar, support load, sales-cycle slowdowns, and brand damage when magic fails. An honest argument that the 'AI-enabled' label should be earned or skipped — not used as decoration.
A reflective essay on being the specialist among generalists — the one person stakeholders ask 'should we use AI for this?'. The leverage, the loneliness, the temptation to evangelize, the burnout that follows, and how I built a remote peer network when there wasn't a local one to lean on.
200K-token windows feel infinite until you measure them. Context is a real budget you spend on every call — every token costs money, competes for attention, and risks pulling the model toward the wrong thread. A practical guide to context efficiency, pruning, and choosing what belongs in the prompt at all.
AI added a third option to the classic build-vs-buy decision: wrap a foundation model API. A practical framework for what to wrap, what to buy, and what to build — and when 'just wrap GPT' is right or dangerously wrong.
A few things I've learned building a business-specialized LLM — why fluency isn't judgment, why structure beats raw scale for domain reasoning, and why grounding every claim in evidence matters more than sounding smart. The methodology, not the secret sauce.
Showing 1-10 of 72 articles