Skip to content

AI / LLM ENGINEERING · PROVIDENCE, RI

Ten years making platforms reliable. Now I'm shipping the AI on top.

Available now for senior AI/LLM roles. I spent 10 years at Modo Labs building the admin and configuration experience for an LLM-powered knowledge assistant, retrieval, indexing, streaming, and guardrails, on a PHP and Kubernetes platform that had to stay up at campus scale.

Résumé (PDF) See the work ↓ Email me →

10+ yrsshipping one campus-scale platform

3case studies, in depth

11 + 2posts and tools on LLM reliability

Anthony Tadros — AI/LLM engineer · available nowTen years making platforms reliable. Now I'm shipping the AI on top: an LLM knowledge assistant with retrieval, indexing, streaming, and guardrails. Available now for senior AI/LLM roles, remote or Boston/Cambridge hybrid. Providence, RI.Work · Case studies · Experience · Tools · Writing · About · Contact · RésuméWriting

I broke my own prompt-caching rule and paid 24 seconds a request — Prompt caching matches an exact prefix. I published that rule in June, then put a clock line in position two of my own system prompt in July, above 4,800 tokens that got reprocessed every time the minute rolled over.

Getting reliable JSON out of an LLM in PHP — Schema-valid and correct are not the same claim. A model can hand back a perfectly shaped object full of wrong values, so the schema goes on the Guzzle call and the value checks go after the decode, before anything else touches it.

How to tell if a prompt change helped — An average is where a regression goes to hide. The new prompt can win overall and still be worse on the category you actually care about, so I run both over a fixed set of real questions and read the per-category difference before the edit ships.

Cost per request, logged next to latency — The price of a call becomes a field on the request, written when the call returns, sitting beside latency and status code. Once it is there, "which feature spent the money" is one group-by away.

Prompt injection in RAG starts with the documents you indexed — Most of the defense belongs in retrieval code, before anything reaches the prompt. By the time a planted sentence is in the context window, you are arguing with the model about whose instructions to follow.

Calling an LLM API from PHP without taking the page down — Nearly everything written about running LLMs in production assumes Python. This is the PHP and Kubernetes version: where the timeout number comes from, and what the page does while the provider is having a bad day.

A 200 you can't use is the new LLM failure mode — The failure with no precedent is a response that arrives fast, parses cleanly, and still cannot be used. Everything else about shipping an LLM feature is remote-call plumbing I already had, so this post spends its time on output validation.

Evals before vibes — Vibes are a measurement, just a bad one: one example, picked by the person who wants the change to work. This is the golden set, the diff against yesterday's run, and the place where LLM-as-judge quietly lies to you.

RAG over a decade of internal docs — Chunking decides what retrieval can find, and freshness decides whether what it finds is still true. Both are settled before anyone types a question.

Prompt injection is just untrusted input — Before the AI work I spent two years pulling Social Security numbers off university computers. A file is bytes until something decides to act on it, which is the whole of prompt injection and most of what appsec already taught me.

Structured outputs deleted half my parsing code — For three years I maintained the layer that turned model text into something PHP could act on. Provider-enforced schemas deleted most of that code, though not the validation that mattered.

Where the money goes at campus scale — Tokens are the row count and the context window is a scan. Every lever I have on the bill is a query-tuning habit from my Oracle days, starting with not paying to re-read the same bytes on every request.