Thomas Creavin

I'm a Machine Learning Engineer at Adobe Research where I work on MLOps. I recently completed my M.Sc. in Computer Science at ETH Zürich, specialising in compiler- and architecture-level optimisation for CPU, GPU, and high-performance computing platforms, alongside coursework in machine learning systems.

My research interests lie in GPU optimisation for scientific workloads and efficient low-precision arithmetic on modern accelerators. At ETH, I was directly supervised by Prof. Dr. Torsten Hoefler (2024 ACM Prize in Computing winner), working on stochastic rounding methods for GPU computing that maintain numerical accuracy while maximising performance.

Professionally, I've worked on LLM observability, MCPs, and multi-agent systems at Adobe Research, scalable data pipelines at Roche, low-latency network infrastructure at Jump Trading, and mail routing systems handling billions of messages at AWS.

Email  ::  CV  ::  LinkedIn  ::  Github

Explored how high-precision calculations can be performed in low-precision on GPU with minimal increase in error. This work addresses a critical challenge in GPU computing where memory bandwidth limitations necessitate lower precision arithmetic while maintaining numerical accuracy.

Developed highly optimised kernels for PolyBench benchmark problems using low-level CUDA API calls in MIT's Julia language, achieving NumPy/PyTorch-like performance. Demonstrated Julia's potential for high-performance GPU computing in scientific applications.

Designed and deployed LLM observability for an Autogen multi-agent platform using Langfuse, DeepEval, and a custom E2E testing framework. Created prompt workflows enabling internal users to request arbitrary data through agents that retrieve schemas from a CosmosDB vector datastore and generate SQL queries.

Built custom end-to-end data pipeline for large molecule experiment data, now deployed to hundreds of lab devices.

Built and deployed Netflow collector in Go for extremely high-throughput network data storage in Clickhouse. Created Ansible host health-monitoring library for safer deployments to heterogeneous crypto clusters.

Contracted to consult and build AWS cloud solution to securely orchestrate IoT devices over OPC UA. The successful prototype is patent pending.

Contributed to the Simple Email Service's mail transfer agent re-architecture, increasing email sending throughput by 10x. Created monitoring systems for CodeDeploy and EC2 bootstrap processes. Recognised with SES Ops Win award for EC2 monitoring contribution and performed on-call duties.

Implemented a highly-optimised version of the seminal Position Based Dynamics paper, achieving a 100x speed-up over naive implementation using C, AVX intrinsics, and a comprehensive Python benchmarking suite.

Created a fast serverless web application for volunteer rostering, treating it as an instance of the nurse scheduling problem with unique constraints. Developed a Gurobi MIP formulation and used supervised search-space pruning to compute near-optimal solutions 92% faster than baseline approaches.

Developed the first-ever implementation of a provably deployment-efficient RL agent, trained on both discrete (Frozen Lake) and continuous (Cart Pole) environments using Python and Gymnasium.

John Kelly Memorial Medal - Top undergraduate graduate (rank 1/115), UCD

UCD Entrance Scholarship - Outstanding Leaving Certificate results

UCD Excellence in Mentoring Award - Peer mentoring contributions

Society Event of the Year - SISTEM 2020 conference organisation

Pat Scanhill Medal - Highest Leaving Certificate result (98th percentile nationally)

AWS Certified Cloud Practitioner - Professional certification

Black Belt in Shotokan Karate - Martial arts achievement

Design and source code from Jon Barron's website