Alex Chaloner
MACHINE LEARNING ENGINEER
ML engineer at Meta working on LLMs for compromise detection at scale - finetuning with GRPO reinforcement learning, ML at internet scale.
Oxford Master's in Mathematics & Computer Science with dissertation in reinforcement learning, and independent projects on LLM pre- and post-training.
Experience
July 2024 - Present
Meta
Software Engineer (Machine Learning)
Machine learning for account compromise
- Led ML within the team, growing ML-based detection from 12% to 70% of abusive-compromise detections.
- Finetuned LLMs with GRPO reinforcement learning for detection and measurement, increasing users detected/measured by LLM systems by 10%.
- Deployed unsupervised ML models at internet scale, protecting millions of compromised users per year.
- Onboarded software engineers onto ML model development, adding protection for millions more users per year.
- Authored the org's scaling strategy: demonstrated empirically that scaling laws held in our domain from 5M- to 5B-parameter models and equivalent data range, determined scaling law formulae, and drove adoption of self-supervised methods to exploit data scale.
November 2021 - July 2024
Ocado Technology
Software Engineer
- Built full-stack software for the automated freezer warehouse in Luton (live since Sept 2023): Java/Spring microservices with AWS SQS, S3 and Lambda; Terraform deployments via GitLab CI.
- Joined the internal ML team and built a GPT-4-powered knowledge-sharing tool (via RAG) for engineers.
- Built warehouse analytics dashboards: Java + Scala services on DynamoDB, with a TypeScript+React frontend.
October 2020 - July 2021
Agari Data, Inc
Data Scientist
- Applied LLMs to detection of malicious business emails; built the internal fraud-analytics frontend in Python.
- Data analysis with heavy SQL usage across large email datasets.
Independent Projects
Muon vs Adam - github.com/AlexChaloner/muon_vs_adam
Empirical comparison of the Muon optimizer against AdamW on transformer training. Muon consistently beats Adam and AdamW in both validation loss and wallclock time, and much less sensitive to hyperparameters.
Empirical comparison of the Muon optimizer against AdamW on transformer training. Muon consistently beats Adam and AdamW in both validation loss and wallclock time, and much less sensitive to hyperparameters.
Self-Distillation Experiments - github.com/AlexChaloner/simple_self_distillation_experiment
Minimal reproduction of Apple's simple self-distillation paper. Replicated the changed top-p shape but could not replicate success - got consistently worse performance on coding benchmark - also designed a closed-form cheap transformation on token distribution to achieve same end.
Minimal reproduction of Apple's simple self-distillation paper. Replicated the changed top-p shape but could not replicate success - got consistently worse performance on coding benchmark - also designed a closed-form cheap transformation on token distribution to achieve same end.
Education
2016 - 2020
University of Oxford
Master of Mathematics and Computer Science (MMathCompSci)
2:1 (upper second-class honours). Dissertation: Hierarchical Reinforcement Learning - compressing action policies via graph abstraction.