Most of my work is reproduction under a strict evidence gate: a claim doesn't count until the number traces back to a file and survives an adversarial check. Currently working through inference-time kernels, quantisation, and torch.compile internals.
BuildFromScratch — single-file LM reproductions at bit-exact parity
SmolLM2-135M and Qwen3-0.6B-Base rebuilt as single-file PyTorch models, verified against the reference weights, then used as a controlled base for architecture ablations at matched compute.
| Check | Result |
|---|---|
| SmolLM2-135M parity vs HuggingFace | max|Δlogits| = 0.0 |
| Qwen3-0.6B-Base parity vs HuggingFace | max|Δlogits| = 0.0 |
| Qwen3 faithful baseline | ~2.14× reference, ~275,000× less training data |
| Modernised IMU-1 bundle @ matched compute | +18% over the faithful baseline (23.52 vs 28.65) |
Three builds — faithful, modernised, exploratory partial-RoPE — each with parity scripts, training pipelines and an lm-evaluation-harness wrapper.
KaggleLab — paper reproductions behind a traceability gate
Small-scale reproductions published on a fixed cadence, with three rules that don't bend:
- Traceability — every published number is parsed from a file in an attached evidence dataset. No number without a source.
- Adversarial re-check — every headline claim gets an independent pass that tries to break it.
- Visible scorecard — each post states its claim class, seeds, comparability, eval status, and what it does not show.
multi-agent-patterns — orchestration for long-running agent systems
Patterns for agent systems that run for hours rather than turns: JobPulse daily automation and MindGraph knowledge visualisation as the working examples.
Patches to the libraries I use daily — kernels, quantisation dispatch, and compiler internals.
| Project | Contribution | Status |
|---|---|---|
| pytorch/pytorch | [dynamo] Split FrozensetVariable off SetVariable to match CPython semantics (+421 / −279) |
In review |
| linkedin/Liger-Kernel | Fix FLCE torch.compile failure from the aten.addmm.dtype_out overload |
In review |
| bitsandbytes-foundation/bitsandbytes | Add an sm121 (GB10) tier to the 4-bit GEMM dispatch heuristic | In review |
Also built
| Project | What it is |
|---|---|
| LetsFineTune | Fine-tuned LMs with LoRA, Unsloth and Transformers — configs, scripts, evals |
| Velox_AI | Voice-agent platform for real phone calls: Twilio + Gemini + Deepgram, hybrid RAG |
| Prism | Research agent decomposing million-token corpora into verifiable reasoning chains |
| 90-days-of-ml | 30+ projects across core ML, MLOps on AWS, and production capstones |
| Deep-Learning-for-Facial-3D-Reconstruction | MSc project: PCA normalisation, 1,200-image rotation sweep, SixDRepNet pose estimation |
ML engineer in Edinburgh. I reproduce language models bit-exact, then run the ablations the papers skip.
Most of my work is reproduction under a strict evidence gate: a claim doesn't count until the number traces back to a file and survives an adversarial check. Currently working through inference-time kernels, quantisation, and torch.compile internals.
- BuildFromScratch — single-file PyTorch reproductions of SmolLM2-135M and Qwen3-0.6B-Base, both at
max|Δlogits| = 0.0against the reference weights. Includes a three-way architecture study (faithful / modernised / partial-RoPE) trained at matched compute. - KaggleLab — small-scale paper reproductions with a blocking traceability gate: every published number is parsed from an attached evidence dataset, every headline claim gets an independent adversarial pass.
- multi-agent-patterns — orchestration patterns for long-running agent systems.
Patches to the libraries I use daily — kernels, quantisation dispatch, and compiler internals.
| Project | Contribution | Status |
|---|---|---|
| pytorch/pytorch | [dynamo] Split FrozensetVariable off SetVariable to match CPython semantics (+421 / −279) |
In review |
| linkedin/Liger-Kernel | Fix FLCE torch.compile failure from the aten.addmm.dtype_out overload |
In review |
| bitsandbytes-foundation/bitsandbytes | Add an sm121 (GB10) tier to the 4-bit GEMM dispatch heuristic | In review |
PyTorch · Triton · CUDA · torch.compile / Dynamo · Transformers · TRL · LoRA / Unsloth · lm-evaluation-harness · TypeScript