Machine learning engineer and independent researcher working on world models and representation learning: what a model actually represents, and whether it has learned the structure of a problem or only the surface cues that correlate with it.
My current work probes predictive world models, using probing and causal intervention to test whether they keep physical state they can no longer see, or only predictive surface statistics (Reverie). The broader programme asks what changes in a model’s representations when you constrain how it learns, through its update rule, its architecture or its objective.
Alongside this: first author on CulturalRiddles, a 51-language benchmark with Cohere Labs, and 2nd place in the Adaption Labs AutoScientist Challenge. Background in electronic and computer engineering, and a decade of forecasting in real operations.
Independent research · Pre-registered
Probing and activation patching of an RSSM world model trained on simulated physics video with no physics in the loss. Tests whether an occluded object’s position stays decodable from the latent state and whether that representation is causally used, with hidden collisions as the central control and CLEVRER for validation.
Cohere Labs Open Science Community · First author
Living cultural-reasoning benchmark: 51 languages, 61+ communities, about 6,100 riddles, 100+ contributors. To be presented at MRL 2026, being submitted to TACL, arXiv preprint coming soon.
2nd place, Data & Visualization track · $1,000 prize
Adaption Charts P2: synthetic multimodal dataset for chart QA and chart-to-table extraction, built so difficulty comes from reading the chart (truncated axes, missing labels, varied visual styles) rather than from arithmetic. Paired Llama-3.2-3B LoRA adapter for table-grounded QA.
Cohere Expedition Hackathon · Pipeline lead, 7-person team, one week
Fully offline document QA for newcomers and migrants; nothing leaves the user’s machine. PDF extraction, BGE-M3 embeddings, ChromaDB retrieval, Aya generation and mDeBERTa hallucination checks. 9,000+ automated evaluations across Chinese, Hindi and Polish showed cross-lingual embedding quality, not the LLM, was the main retrieval bottleneck.
Adaption Labs Uncharted Data Challenge · Grade A (9.4/10)
187 curated Q&A records across 20 categories and 48 crops, built from public-sector sources over ten iterative submissions. Honorary recognition from Sara Hooker. Follow-up: an 11-language Indian agricultural dataset.
Plus six public datasets and a LoRA adapter on Hugging Face (580+ downloads).
Cohere Labs Open Science Community · Remote
The Meta Gradient · Toronto (Remote)
M2M Tech · Vancouver (Remote)
Whole Foods Market · Toronto
University of Windsor
Thesis: FPGA implementation of face recognition using Eigenfaces
Anna University
University of Toronto