I’m a graduate researcher at the University of Toronto, working in the Data-Driven Decision Making Lab (D3M). My research is neuro-symbolic architectures that can verifiably reason and plan in open-world environments, with detours into interpretability, RL, and continual learning along the way.
Current work
I’m currently building autonomous neuro-symbolic RL agents that discover the mechanics of unknown environments through systematic experimentation, knowledge representation, and world-model learning.
My first publication was accepted to ICLR 2026 (Poster). We introduce a neurosymbolic planning framework that brings symbolic guarantees to LLM agents in partially-observable environments. It improves on previous baselines by 373% and attains state-of-the-art across central benchmarks.
See: Natural Language PDDL (NL-PDDL)
Research interests
- Reinforcement learning
- World model building
- Mechanistic interpretability / XAI
- Verifiable formal reasoning
- Continual learning
- NLP and CV in neural space
Experience
- Contributing author on an ICLR 2026 paper on neurosymbolic agentic planning for LLM agents — 373% over baselines, SOTA across benchmarks.
- Building autonomous neuro-symbolic RL agents that discover open-world mechanics through experimentation and world-model learning.
- Invented sparse-autoencoder dictionary learning to extract personality traits from LLM latents — steers behaviour via neuron activation instead of prompting.
- Found universal features across LLM architectures; 2+ orders of magnitude more efficient feature evaluation.
- Shipped a multi-stage surgical block-schedule optimizer to production — ~C$300M in added revenue across three Ontario hospitals.
- EDAs drove pitches to EY, KPMG, and the OHA Board, securing federal approval for Canadian hospital rollout.
- Refactored 50k+ lines of React to ship LoRA fine-tuning and distributed inference for LLM/VLM/VLA/Diffusion/CNN via ONNX.
- Built a 4 PB PostgreSQL backend over 1000+ cores; client demos contributed to raising $1M+ in VC funding.
- Stacked Gated SAE architecture for higher-dim latent extraction in PyTorch.
- Led 4 PMs and 15+ engineers across 8 months; onboarded members on Git, research methods, and conference-deadline development.
- Density-based CNNs for U. Nairobi maize-yield forecasting — 5+ orders of magnitude labour reduction. CUCAI23 Sustainable Tech award.
- Reimplemented and extended a tiny-object detection paper in PyTorch for non-local yield and disease detection.
- Brought Queen’s CSA website in-house — cut hosting and dev costs by 90% while serving 1800+ monthly students.
- Trained 3 first-year devs; introduced agile methodology and contingency planning, cutting critical-path time by 20%.
- Automated employee-portal tests with Selenium + Oracle Testing Suite — 98% reduction in test time.
- Integrated Oracle-compatible libraries into the CI/CD pipeline; revamped UI test traceability.
Education
- GPA 4.0 / 4.0 — LLM Reasoning, Embodied AI, Reinforcement Learning.
- GPA 3.95 / 4.0 — Mathematics and Artificial Intelligence Specialization.