I work the whole path an AI feature takes to get real: retrieval that finds the right evidence, a model tuned and served inside a latency and cost budget, guardrails that keep it from asserting things it can't support, and an eval harness that decides whether any of it shipped.
The demos are easy. What I care about is whether the thing holds — so my projects carry baselines they could lose to, confidence intervals, cost per request, and a written verdict when the answer was no.
Retrieval, agents, fine-tuning, serving, and the measurement around them. Every one runs on real data, states what was measured and what wasn't, and links to its full source. Every card flips — the front is the claim, the back is the evidence.
Clinical notes to ICD-10-CM codes, retrieved and reranked at production scale.
Hybrid BM25 + MedCPT retrieval over 72,750 codes, cross-encoder reranking, and a QLoRA-tuned Llama-3.2 served on vLLM behind a semantic cache. Scored on 1,000 held-out notes.
Three retrieval arms over one corpus, measured with confidence intervals.
Baseline, graph-RAG and agentic arms scored on one corpus with bootstrap CIs. Agentic retrieval nearly doubles recall — and answer F1 barely moves, so retrieval was never the constraint.
Six agents plan a real trip without ever being trusted with a fact.
Six agents with narrow tool sets each. Every price is checked against the payload it came from, all arithmetic is Python, and uncertainty is labelled — on free-tier APIs throughout.
A coding agent that has to prove its patch works before it can finish.
Writes a failing acceptance test first, patches, then verifies through deterministic gates: digest-checked tests, static checks before any model review, bounded repair, enforced budgets.
Job postings turned into a queryable company and tech-stack signal.
Scraping, rule-based labeling, a QLoRA-tuned extractor served on vLLM, a DuckDB warehouse, and a Marimo dashboard on Cloud Run.
PDF chapters into summaries, flashcards, a glossary, and topic clusters.
gpt-4o-mini paired with classical NLP — KeyBERT and spaCy for terms, UMAP + HDBSCAN for clusters — and graded with ROUGE against ground truth.
Tools I've used to ship something on this page — not a list of everything I've read about.
An LLM feature that works in a notebook is the easy 20%. The questions I find more interesting come after it: does the retrieval surface the right evidence on inputs nobody curated, can the system tell you when it doesn't know, what does a request cost, and how would you notice if it quietly got worse?
That shows up throughout this work. A semantic cache and a cross-encoder make a clinical retrieval system fast and accurate enough to trust. A multi-agent planner checks every price against the tool payload it came from and does its arithmetic in Python, because a model that mis-adds a budget is worse than no budget at all. A coding agent has to produce an observed failing test and a passing containerized run before it's allowed to call itself done.
I also report what didn't work. One project retracts its own earlier retrieval finding once the ablation was run properly — the corrected result is indistinguishable from zero, and it sits in the README next to the wins. Another measures a near-doubling of retrieval recall and says plainly, in the same table, that answer F1 barely moved. A result you'd bury when it disagrees with you isn't a measurement.
I'm comfortable across the whole path — retrieval and indexing, fine-tuning, the serving layer, the eval harness, and the tracing and cost accounting that keep it honest — on AWS, GCP and Azure.
I'm looking for AI Engineer roles. The fastest way to reach me is email.