Jonathan Cassens

1. Now

   ML Engineer, Kpler

   Commodities tribe

   • Led the SageMaker evaluation end to end, through to reporting: training and serving models via MLflow, Studio spaces, image dependencies, and role and permission scoping.
   • Validated data access on the new CDC and dbt pipelines through Redshift.
   • Contributing to the move off Dataiku to our own ML platform and kpler CLI.
   • Improving the EBX, EBD and EBA models through feature engineering and model design.

   SageMaker · MLflow · Redshift · dbt · Kafka · S3 · Iceberg · IAM · PostgreSQL

2. 2026

   AI Engineering Intern, Medly AI

   • Production pipelines for an AI tutor with 300k+ users.
   • A question difficulty transformer driving personalised question assignment.
   • An LLM student simulator and eval framework, plus SummaryEval, an LLM-judge quality gate.

3. Feb to Jun 2026

   GenAI Alignment Engineer, Google via QualityAI

   • Physics data curation for LLM pedagogy alignment.
   • Failure-report triage pipeline: routing over 91% faster.

4. Summer 2023

   Software Development Placement, Sky

   • React component refresh that cut content tagging time by 17%.

• Medly AI · Transformers · IRT

  Question difficulty transformer

  Predicts question difficulty so each student gets one pitched at the right level.

  ~90% in bracket · 99%+ within one

• Trading · Game theory

  IMC Prosperity 04

  Market making, ETF arbitrage and game-theoretic allocation.

  Top 1% manual · top 2% algo

• Hardware-aware ML

  Wildfire detection on an FPGA

  A pruned, INT4-quantised CNN deployed through ONNX and HLS4ML.

  405k params · ~94% recall

• LLMs · Evals

  Context Pareto

  Accuracy against tokens for long-context compression strategies.

  Public

• Monte Carlo · Football

  FPL modelling engine

  AI-native Monte Carlo simulation of my Fantasy Premier League team.

  ~500k global rank, climbing

• Reinforcement learning

  Poker Q-learning leak finder

  Exploits opponent tendencies and finds leaks in my own play. 200k+ hands.

  +8 bb/100, simulated data only

• RAG · Local LLMs

  arXiv Physics RAG

  Ask questions of physics papers with Ollama, ChromaDB and React.

  Public

• Edge ML

  Model to Mask

  Taking neural networks and small language models to ASICs. A curiosity project.

  Public

• Tooling

  WhatsApp bot remover

  Spots bots in group chats, deletes their messages and kicks them. Built for uni chats.

  Personal tool

Also on GitHub higgs-boson-classification electron-muon-scattering lstm-double-pendulum pca-mnist covid-19-predictor tenant-matchmaker-api

Platform and data

AWS, SageMaker, IAM, S3, Redshift, dbt, Iceberg, Kafka, PostgreSQL, Docker, HTCondor

ML and AI

PyTorch, TensorFlow, scikit-learn, transformers, GANs, LLMs, RAG, LLM evals, reinforcement learning

MLOps

MLflow, Weights & Biases, Git, CLI tooling, Dataiku (migrating away)

Languages

Python, SQL, JavaScript, React, Java, pandas, NumPy

Maths and physics

Statistics, item response theory, stochastic modelling, quantum information

Hardware-aware ML

FPGAs, HLS4ML, ONNX, pruning, quantisation, HPC

• Poker

  Former pro, then a dealer for a while. The statistics are what stuck.

• Volleyball

  Ten years competitive, still in the London league. Two years as a volunteer coach, two promotions.

• Reading

  AlphaSignal and TLDR to keep up. Standing opinion: DeepSeek is the GOAT, a Pareto legend.

• Food

  My favourite hobby, by some distance. Hence the series below.

• Games

  Chess, board games, video games, and FPL.

• Outdoors

  Travelling and hiking, ideally towards somewhere good to eat.

Happy to talk ML platforms, evals, physics, or where to get lunch.