About
Production AI, where being wrong has a consequence.
I'm Josh Zirena, an AI Tech Lead based in Miami and an FIU grad, with a little over nine years as an engineer. Right now I'm AI Tech Lead on a global pharmaceutical company's internal AI platform, working mainly on its application service: it registers MCP servers, handles their auth and configuration, and routes their traffic through the LLM gateway. Before that I led six engineers at a federal government contractor for three years, building for a federal healthcare agency, including the daily evaluation pipeline for the AI assistant on a federal public health website.
Almost all of that time has been in industries where being wrong has a consequence: open banking at a European fintech first, then healthcare. That shapes how I build AI.
Most AI work fails the same way: an impressive demo that quietly falls apart when it meets real users, real data, and real reliability requirements. I specialize in that gap: the architecture, evaluation, guardrails, and observability that turn a prototype into something a team can actually run.
I care about systems that are grounded (answers can be checked against their sources), measurable (you know when they get better or worse), and maintainable (the team can understand and extend them). I try to leave every team more capable than I found it.
How I build AI
Turn silent failure into loud failure
A deterministic system throws when it breaks. A model returns a confident, well written, wrong answer and nothing in the stack objects. The engineering isn't making the model better; it's making its failures visible.
AI proposes, deterministic code disposes
The model never commits state and is never the system of record. Validation, policy, and the database sit around it. The model reads the question; the data holds the truth.
Measure on a schedule
A fixed set of known good answers, run daily, because content, indexes, and model providers drift even when nobody deploys. Trace everything, because the bad answer looks like a success.
Earn autonomy
Shadow mode, then a human reviews everything, then auto-approve only the high confidence slice, widening it as the data earns it. Anything that moves money or commits a decision: the AI prepares it, a person commits it.