I build reliable AI systems at scale.
An AI and ML platform leader with 5+ years deep in AI systems and two decades building at scale. I've built the ML platforms behind Amazon One — on a foundation of large-scale infrastructure for HSBC, Visa, E*Trade — and scaled teams from 10 to 100+, pairing technical depth with business acumen to drive outcomes at every scale. Now CTPO at Future Family, researching what makes AI agents reliable in production.
The systems I keep coming back to.
In an agent, the scarce resource isn't compute or memory — it's latency, and the external calls that fail. Every step is a slow, fallible round-trip to a model or a tool, so the system's real job is to tolerate failure and never pay twice for the same expensive work. That's where most of my technical attention goes.
Event-journaled runtimes
Every step is a slow, fallible call to a model or a tool. Recording each one means a failure resumes from exactly where it broke — and you never re-pay the latency for work already done.
Relevance is the lever
A model is only as good as what you put in front of it. Getting the right context — and only the right context — to the model is the highest-leverage way to lift quality and cut cost. Most "model" problems are really context problems.
Regression detection
Replay real production runs against each new model version to catch quality regressions before they ever reach users.
Model · Memory · Design
My framework for optimizing inference across accuracy, latency, and cost — three pillars, each tuned against a golden dataset so speed and cost gains never quietly erode quality. Compiling to the right target is one lever: ONNX Runtime for portable CPU and edge, JAX / XLA and GPU graph compilers for high-throughput serving.
- Distributed systems
- ML platforms
- Inference optimization
- Large-model training
- Model compilation
- Context & memory
- Agent runtimes
- Reliability & replay
From fintech infrastructure to ML platforms to leadership.
Twenty-plus years, one throughline: making large, failure-prone systems dependable.
Field notes you can take with you.
Practical guides drawn from the work — free to download and use while you build.
The MMD Playbook
A practical guide to scalable AI inference — the Model–Memory–Design framework, golden datasets, a stage-based playbook, and the KPIs that keep accuracy honest.
ONNX vs JAX/XLA: a decision guide
How to choose a compile target while you're building — CPU vs GPU, portability vs throughput, and when each runtime is the right call.
Notes from inside the machine.
Essays on the systems that make modern AI run — how it trains, how it serves, and how it's composed.
Code in the open.
Tools and experiments I build and share. (Descriptions below are placeholders — worth a one-line tagline each.)
Building something hard?
Let's talk.
Open for conversation — challenging problems, technical deep dives, and complex AI systems.
Or reach me directly —