Mike Malloy
LLM Engineering & AI Systems

Building intelligence that
performs under pressure.

Custom RAG pipelines, fine-tuned models, and production ML systems on AWS Bedrock and SageMaker — battle-tested on enterprise data with measurable results.

From Michelin-starred kitchens to Fortune 500 AI systems — I bring the same obsessive precision to language models that I brought to the pass. The details are where it wins or breaks.

Get in touch View capabilities
Latest Evaluation RUN #4412
servicerag-pipeline
modelclaude-ft · v3
statusrunning
recall@100.00
hallucination0.0%
p95 latency0ms
Bedrock Lambda RAG

Every model ships measured — or it doesn't ship.

6×Retrieval Precision Improvement
80%+LLM Projects Fail in Production
F500Enterprise Experience
About

The same discipline, a different kitchen.

RAG systems that hallucinate under real queries. Fine-tuned models that ace benchmarks but drift in production. Bedrock deployments where nobody measured anything. The same pattern keeps showing up.

I started in technology, then spent years training in Michelin-starred kitchens where precision and flawless execution under pressure weren't optional — they were survival.

That taught me complex systems either work exactly as designed, or they fail at the worst possible moment.

When I returned to tech and went deep into LLM engineering, I found the same gap: teams shipping models that impressed in demos and collapsed on real data. The tooling had changed — Bedrock, SageMaker, vector databases, fine-tuning pipelines — but the discipline around evaluation, monitoring, and production readiness hadn't kept up.

I build RAG pipelines, fine-tuned domain models, and production ML systems on AWS that teams can actually maintain. The goal is always the same: make moving fast and building reliably the same thing — not competing priorities.

Capabilities

What I build.

Scalable AI infrastructure. Measurable performance. Production-ready from day one.

01

Retrieval-Augmented Generation

Pipelines that surface the right information at the right time, at scale — vector-database architecture, embedding optimization, hybrid semantic + keyword retrieval, and disciplined context-window and chunking strategies.

02

Fine-Tuning & Model Adaptation

Domain-adapted models using LoRA and QLoRA, evaluated against real-world benchmarks and edge cases, with curated training data and A/B testing frameworks that prove performance instead of asserting it.

03

AWS Bedrock

Production LLM deployment and infrastructure — prompt engineering, guardrails for safety and compliance, and cost and usage monitoring that keeps spend predictable under real traffic.

04

SageMaker Pipelines

End-to-end ML workflows with full observability — model monitoring, drift detection, automated retraining triggers, and feature stores with data versioning teams can maintain.

Selected Work

Production systems, end to end.

di4health — Public Health Intelligence Assistant

di4health — Public Health Intelligence Assistant

Ask a plain-English question — "Compare obesity rates in Cook County and Harris County," or "What has the CDC reported on recent measles outbreaks?" — and get an evidence-backed answer in seconds. Under the hood, a multi-tool agent decides how to answer: querying CDC PLACES statistics across 3,000+ U.S. counties, pulling CDC WONDER mortality data, or running semantic search over CDC MMWR outbreak reports. It deliberately pairs two retrieval styles — structured SQL over tabular health data and vector RAG over narrative reports — then returns a structured intelligence brief with the numbers, sources, and caveats spelled out, rendered beside a live chat thread. Authenticated and deployed end to end (Next.js + FastAPI agent), every response is grounded in real data and names the tools it used. Built to demonstrate a production-grade agentic RAG system — typed output contracts, fail-fast startup, graceful degradation — not a demo that dazzles once and breaks on the second question.

Agentic RAGVector SearchFastAPI
AI Digital Assistant

AI Digital Assistant

An interactive digital twin powered by Amazon Bedrock and AWS Lambda. Ask questions and learn about my background, experience, and approach to LLM engineering.

BedrockLambdaRAG
Molecular Toxicity Screening Platform

Molecular Toxicity Screening Platform

Paste in a SMILES string and screen a compound against 12 Tox21 toxicity endpoints in seconds — nuclear receptor and stress-response pathways like NR-AR, NR-ER, and SR-p53. A fine-tuned ChemBERTa transformer drives the predictions, paired with an RDKit-powered ADMET profile: Lipinski and Veber drug-likeness rules, PAINS alerts, and key molecular descriptors. Results roll up into a weighted composite risk score with Low/Moderate/High tiers and a rendered 2D structure. Built to demonstrate a production-minded ML workflow — domain model fine-tuning, explainable outputs, and confidence-aware scoring — not to replace the wet lab.

ChemBERTaRDKitADMET
Get in Touch

Let's build something that holds up in production.

Based in

Available worldwide