Work

Hello! I am an ML software engineer and leader. In addition to direct contributions, I have built and led engineering teams, run product management, and served as an executive in the crucible of startup life.

At Google, I led Emerging ML, the team which develops and maintains new ML defenses for Google Ads, and I’ve also been central to LLM adoption related to these defenses. Earlier, I built the distracted driver detection system used by Cruise vehicles during COVID, deployed in 600+ vehicles. Earlier, for four years, I CTO’d Topology Eyewear, a startup which uses iPhones, ML, and CNC milling to make custom-fit eyeglasses. I also have wide consulting experience ranging from startups to F500s.

I am passionate about understanding, extending, and applying the capabilities of LLMs and other forms of Generative AI. (The computers are talking to us—what could be more exciting?!) I love working on teams that care about what they’re doing.

I live in San Francisco but spent years in Cambridge, New York, Milan, Oxford, and London. I enjoy working in diverse teams.

US/EU citizen. BA (Physics), PhD (Mathematical Biology).

The employee roles below were at Google, Topology Eyewear, and the London Institute for Mathematical Sciences. The rest were through my consulting business.

Work History

Answer.AI

• Member of R&D Staff
  April 2024 – Present · San Francisco, CA

  I am working directly on AI research and development. Hooray! These times are way too interesting to miss being right in the action.

Google

• Senior Staff Software Engineer, Google Ads
  October 2021 – March 2024 · Mountain View, CA

  Google Ads is under constant attack from adversaries seeking to defraud advertisers by pretending to be users. But Ads still works because behind the scenes AdSpam, a unit of hundreds of engineers and analysts, maintains an evolving suite of defenses.

  I led the team which develops new ML defenses and maintains existing ones, and the team which measures defense success for senior management. I am not permitted to discuss methods in detail.

  • LLMs and Generative AI. In the LLM Working Group, I was responsible for AdSpam, representing 200+ Googlers, catalyzing 6 LLM projects with XX M$ impact, and presenting these projects to leadership, with my assessment and plan for organizational adoption. Judged and organized content for hackathons introducing 100s of Googlers to LLMs. My team worked on the first use of LLMs in Ads to create a targeted signal pipeline for detecting invalid traffic.
  • ML operations. My team, Emerging ML, ran ML defenses which filter trillions of events/day, with millisecond latencies, protecting billions of dollars of yearly impact. These defenses are critical for Ads: ML or infra failures have immediate, measurable consequences on finances and on partner relationships.
  • ML modeling. I managed my team’s development and launch of 5+ defenses in multiple Ads products (Web, Mobile, and Search), catching XX M$ in invalid traffic. Innovated privacy-safe defenses, achieving 140%+ replacement effectiveness. Work touches on various ML techniques and concerns: DNNs, transfer learning, autoencoders, transformers, quantization, distillation, GNNs and graph modeling, explainability, and different supervision modalities to handle label scarcity.
  • Management. Over 5 Perf/GRAD cycles, for 10+ SW/ML engineers, I managed performance reviews, promotions, hires, and also smoothly managed low performers. Baselined defense velocity, aligned the org on this measure, identifying 3 key bottlenecks. Drove production reproducibility, with baselining and analysis that helped unblock 5+ pipelines that could not otherwise be refreshed. Rebooted my teams’ project process, set up forums to drive engagement with ML outside of Google, and led workshops to align on an overdue pivot.

ObservantAI

• Machine Learning Engineer
  June 2020 – November 2020 · San Francisco, CA

  ObservantAI, a YC startup, provides the safety driver monitoring system for Cruise, a top autonomy startup. I personally built the distracted driver detection model, modeled from scratch to handle COVID-19 face coverings, which deployed to 600+ vehicles.

  • Analysis. I analyzed requirements regarding false negatives vs. positives, benchmarked the existing online-only system to determine its performance before and after face coverings, and defined metrics suitable to this anomaly detection problem with imbalanced classes.
  • Model development. I prepared the data and explored many architectures and training treatments, varying batch size, training time, sampling strategies, use of fine-tuning, label smoothing, and many data augmentations. I achieved <20ms latency on Apple Neural Engine accelerators, delivering CoreML models with well-defined precision/recall tradeoffs, along with Grad-CAM explainability artifacts and a complete development rig.
  • Tooling and deployment engineering. I set up model development and evaluation infrastructure (Ubuntu, Jupyter, fastai, PyTorch, DVC), and did low-level work to overcome platform bugs. Developed benchmarking tools, and CoreML tools to observe intermediate outputs and to perform ad hoc model surgery. Enhanced Apple’s own coremltools to convert from undocumented PyTorch layer types.

Topology Eyewear

• Chief Technology Officer
  December 2015 – March 2020 · San Francisco, CA

  Topology Eyewear is a San Francisco startup which uses machine learning, computer-controlled manufacture, and augmented reality to measure, make, and sell custom-tailored eyewear—glasses made from scratch to fit the customer’s face. I was the CTO for over four years, from when the company did not have an app and had not sold a single pair of glasses, to where my team’s apps had sold thousands of pairs, on the App Store and in the retail shops of our strategic partner, New Look, Canada’s second largest eyewear business.

  • Management. I hired and managed dozens of software engineers in backend development and iOS development. I recruited talent everywhere, locally and overseas, from tech newcomers to Facebook alums. I worked closely with our product designers, also serving as de facto product manager for years. I participated directly in all major strategic discussions and in fundraising. I also dealt with the gritty parts of startup life that are not glamorous to talk about. I earned and kept the trust of my colleagues through ups and downs and then ups again.
  • Consumer and retail iOS apps. I built major parts of our consumer iOS app, a high-polish app with 40+ screens to handle interactive face scanning, an augmented reality virtual try-on view, product navigation and details, optimized onboarding flow, payment capture, order fulfillment communication, analytics, etc. My favorite components, such as the face scan and virtual try on, required combining an awareness of user-centric design with deep technical work at the limits of the platform’s capabilities. Later I led the critical and accelerated effort to build the retail iPad app, which integrated on-device face reconstruction technology and worked in an unusual, long-running deployment environment.
  • Machine learning infrastructure. I productized our custom 3D face reconstruction ML models which power the in-app augmented reality experience. I built the system that provided data control, versioning, reproducible deployment, and monitoring, well before “MLOps” became a buzzword and a turnkey product. I designed and implemented the scalable, high-uptime infrastructure for running these models reliably, using Clojure, Datomic, bash, and AWS services to build a scalable distributed task queue with analytics and admin functionality. This system ran for 3+ years with only one unplanned downtime event, and smoothly handled 8× traffic spikes during media showings.
  • System architecture. I selected and integrated the stack of third-party support services which allowed us, with only a couple of engineers at first, to run a custom retail app, taking payments, managing customer service and fulfillment communications over web, email, and app touch points, integrating with a bespoke factory. I am proud of the many systems I designed to avoid building.

Selected Clients

• McKinsey & Co.
  2011 – 2013 · London, UK
  Led the iOS engineering effort and contributed to the UX of two critical sales support apps, used by management in pitches to Fortune 500 clients. Designed and built animations, simulated physics, a CoreText text system, and a pipeline to prepare and incorporate hundreds of image, media, and web assets. Harmonized text encodings, time zones, and personalities. This work was highlighted as exemplary by CTO John Anderson and has been forked to create a half dozen variants used by McKinsey’s global offices.
• Deutsche Bank
  2009 – 2011 · London, UK
  Worked with traders to enhance the trading system driving the world’s largest foreign exchange options desk. Added new convertibility risk metrics, such as fxgamma, for Brazilian FX swaps, gathering requirements, applying options pricing theory, and safely modifying a large, uncommented legacy C++ codebase.
• Canadian Society for Aesthetic Plastic Surgery
  San Francisco, CA
  Designed and built the iOS app and Clojure backend used by hundreds of surgeons for viewing streamed, encrypted surgical videos on iPhones and iPads. Safeguards patient information, blocks credential sharing, adapts to bandwidth conditions, supports AirPlay, works in French and English, and facilitates administration.
• Dalmine Energie
  Milan, Italy
  Taught an in-house course on Java and object-oriented programming (in Italian) to the team of Visual Basic developers and led development of an internal workflow tool.
• London Institute for Mathematical Sciences
  London, UK
  Postdoctoral research in mathematical models of evolutionary theory.

Education

• Ph.D. Mathematical Biology, Oxford University
  Oxford, UK
  My PhD thesis, Evolvability: a Formal Approach, developed a novel, mathematically formal approach to describing evolvability—connecting classical evolutionary theory, modern evo-devo, and evolutionary computation. I also invented genospace algebra, a Herbrand logic with ground terms in graph theory, to analyze the mutational structure of genotype space.
• B.A. Physics, Harvard University
  Cambridge, MA
  Completed undergraduate requirements early and took graduate courses in quantum mechanics and quantum field theory.
• Online Courses
  Convolutional Neural Networks ◇ Structuring Machine Learning Projects ◇ Improving Deep Neural Networks ◇ Neural Networks and Deep Learning ◇ Machine Learning (Stanford) ◇ Programming Languages (University of Washington) ◇ Introduction to Logic (Stanford)

Books

• Worked closely with Jeremy Howard and Sylvain Gugger of FastAI, helping with editing and some writing in Deep Learning for Coders, with Fastai and PyTorch.
• Machine Learning by Tutorials, 2nd edition — I updated and rewrote the chapters on natural language processing and sequence-to-sequence models for this edition.
• The Swift Apprentice, later editions — I wrote the chapters on generics and value semantics.
• The Swift Apprentice, 1st edition — I wrote the chapters on generics and functional programming.
• Technical editor for more than forty-five Swift and iOS development articles on RayWenderlich.com, improving the flow of explanation and verifying all code and technical information.

Languages

I have worked professionally in Swift, Objective-C, C++, C, Java, JavaScript, SQL, Python, Ruby, Mathematica, Clojure, Common Lisp, and bash. I have studied Scheme and Standard ML. Some French, Italian, Modern Greek.

Theater

Theater? Well, since you ask… Yes!

For about 15 years I studied, taught, and performed Chicago-style longform improv comedy. I studied with Paul Sills (Second City), Armando Diaz (iO), Mick Napier (Annoyance), Ian Roberts (UCB), Shira Piven (Burn Manhattan), Andrew Secunda, Todd Stashwick, Jay Rhoderick (Centralia), among many others. I have taught classes and workshops at the introductory and advanced level and conceived and directed shows.

I have performed in shows ranging from plain vanilla Harolds, to improvised musicals, news spoofs, and (my favorite) hourlong two-handers in the naturalistic style that TJ & Dave popularized years ago.

What does this have to do with technology? Very little! Except, of course, it tells you something about me, and it informs some of my strengths as a team member and problem solver. I listen deeply and collaborate well. I am very comfortable with uncertainty and exploration that is part of research and product development, not just the more well-defined, execution-oriented stages of such work. I value originality, creativity, and self-expression. I am in my element presenting work to other people and collaborating in workshop environments. I like working with others.

Two-Page PDF

Too much information? Here’s a two-page resume.