AI engineer in London. I build agents and the benchmarks that keep them honest: for wet-lab automation, GPU/TPU kernels, hardware design and protein models.
A benchmark that tests whether AI agents can turn the methods sections of published papers into executable liquid-handling robot protocols, aimed at science's reproducibility crisis.
A fast fire-spread simulator and RL sandbox, built in 6 hours. An air-tanker agent learns where to drop water. The trained agent saves about 98% of fuel, against 42% with no intervention. Judged by researchers from NVIDIA, Project Prometheus and NASA.
Distilled Poolside's 33B, 256-expert Laguna XS.2 MoE into a 3B fully dense model. Then SFT and GRPO on the Sakana AI CUDA Engineer archive for KernelBench-style kernel optimisation, with a vLLM engine for fast pass@k evaluation.
Added Verilog/SystemVerilog support to Sakana AI's evolutionary program search, gated by Yosys SAT-equivalence checking. Evolved designs beat the human reference on 27 of 45 RTLLM v2.0 designs for area, depth and power. The best result scored 175 against the reference's 100.
The first benchmark for LLM-generated JAX Pallas kernels on TPU and GPU. Three difficulty levels, correctness-gated speedup scoring against JAX baselines, robustness filters, CI and a Hugging Face dataset.
KernelBench, except the thing being optimised is JAX itself. Agents rewrite C++/CUDA hot paths and XLA passes, JAX is rebuilt, and the score is the speedup of the real library, gated on correctness.
A ProteinGym-compatible benchmark for zero-shot variant effect prediction across the Parkinson's gene panel (SNCA, LRRK2, GBA, PRKN, PINK1…). Finding: ESM-2's signal inverts on SNCA, and larger models make it worse.
Reproduces the published ProteinGym baseline for ESM-2 650M (ρ = 0.414) on Modal A100s, pinned to exact SHAs. Compares masked-marginal, batched and single-pass scoring, followed by kernel-level work in Mojo, CuteDSL, Pallas and quantisation.
Ten parallel ShinkaEvolve instances attack open Erdős problems, with Lean 4 proof verification and a live constellation UI. Built at the {Tech: Europe} London AI Hackathon.
A 3D MuJoCo renderer for Opentrons OT-2 protocols, driven by opentrons_simulate logs. It powers the Text2WetLab videos.
Slope-coupled fire spread on heightmaps, and voxel buildings where fire fills rooms and crosses doorways, scored by an eval harness.
Intrinsic's AI for Industry Challenge: cable-insertion policies in MuJoCo, Isaac Lab, Gazebo and ROS 2, with SO-101 tooling.
Top of the postgraduate cohort. Autonomous navigation following AprilTags, plus fine-tuned YOLOv8 bottle detection in Nav2 and Isaac Sim.
Coursework and experiments for the SO-101 robot-learning course on LeRobot.
Per-design PPA results, reward-hack audits and figures for the evolved Verilog work.
Expert-activation analysis across code datasets that informed the MoE → dense densification.
Merged Verilog support in Sakana AI's ShinkaEvolve and a Gymnasium environment listing, plus working forks of JAX, PyTorch, ESM, openevolve and LeRobot.
A UK scam-call honeypot. AI voice personas waste scammers' time, and the transcripts are categorised to publish trends that protect vulnerable people.
Real-time video risk detection for suicide prevention. Won Tech for Good at BathHack 2022 and was later extended with suicidology researchers.
A voice agent and carer dashboard for elderly medication reminders. Won 2 prizes at the AI Engine Hackathon (Best Use of ACI.dev, Largest System Built).
Hardest long-video samples per category, used for Emenate's VLM research.