About

Hi, I'm Benjamin

Life long learner and problem solver.

I am a Research Scientist at InstaDeep with a PhD in Electronic Engineering, working at the intersection of machine learning research and engineering. My current focus is deep learning for genomics, where I build distributed training workflows and high-performance data infrastructure for large-scale experiments.

My background spans reinforcement learning, robotics, and deep learning, and across those areas I have been most drawn to the challenge of building systems that are both technically strong and practically reliable. Recently, that has included post-training a 650M-parameter model, designing pipelines for 30TB datasets, and supporting multi-day distributed training on H100 clusters.

I am now looking to bring that experience more directly into production machine learning engineering, applying a research foundation to building robust, scalable systems that make advanced models genuinely useful in real-world settings.

Outside work, I am shaped by my Christian faith and enjoy family life, thoughtful conversation, and time in the garden.

Professional history

Experience

Research Scientist (machine learning)

– present

InstaDeep · London, UK (hybrid)

  • Developed post-training for the NTv3 model predicting experimental data from DNA; achieved state-of-the-art performance on genomic annotations and functional track prediction.
  • Built efficient dataloading for 16k experiments across 24 species (~30 TB): TensorStore-backed remote storage and compression, averaging ~800 MB/s read throughput during training.
  • Ran large-scale distributed training of a 650M-parameter U-Net (hybrid CNN + Transformer) on 8× H100 GPUs for 10 days with production-grade, fault-tolerant code.
  • Significant engineering contributor: over 90k lines of code across 140+ pull requests in the last year.
  • Exposure to agentic systems, LLM finetuning, RAG, and multi-modal models.

Postdoctoral researcher

–

Stellenbosch University · Stellenbosch, South Africa

  • Research at the intersection of machine learning and analytical modelling: 3 journal articles (IEEE, Elsevier), 4 IEEE conference papers, and a survey — collectively 160+ citations (Google Scholar).
  • Compared SAC, TD3, and DDPG with varying inputs for autonomous racing; results supported deployment of an end-to-end SAC agent on a physical vehicle at up to 5 m/s.

Education

PhD, Electronic Engineering

Stellenbosch University

  • End-to-end reinforcement learning for autonomous racing with TD3; novel trajectory-based reward signal improved completion rate by 300%, with peak speed 8 m/s.
  • Applied viability theory so agents could train without crashing: safety supervisor improved sample efficiency 5×, zero crashes in training, validated on a physical robot.

BEng (Mechatronics)

Stellenbosch University

  • Coursework included engineering mathematics, applied mathematics, C programming, engineering design, and thermodynamics.

Projects

Projects I have worked on at various points in my life.

Coding Projects

F1Tenth Racing Benchmarks

A project to benchmark common algorithms for F1Tenth autonomous racing. I implemented trajectory optimisation and tracking, model predictive control, and end-to-end reinforcement learning methods.

PPO in JAX

The proximal policy optimisation algorithm using the JAX framework. I developed this project to learn how the JAX ecosystem can be used for high-performance code.

Deep reinforcement learning algorithm implementations

Personal implementations of the DQN, PPO, DDPG, A2C, SAC and TD3 algorithms in PyTorch. The repo includes a document summarising the algorithms and code components.

Sensor fusion algorithms

Implementations of the Linear Kalman Filter (LKF), Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF) and Particle Filter (PF)

Tutorial for end-to-end reinforcement learning for autonomous racing

A tutorial style example of how to use end-to-end reinforcement learning (TD3 and SAC algorithms) for F1Tenth autonomous racing.

Publications