Introduces the Frequency Synchronization Degree (FSD), a permutation-tested metric for Fourier circuit synchronisation that requires no prior circuit knowledge. FSD predicts grokking 500-3,000 steps in advance across nine modular-addition configs (mean lead +1,722 steps), synchronising before a restricted-logit baseline in all nine. Weight-decay intervention at the FSD ceiling causally confirms the inter-phase gap is regularisation: ∆t ∝ 1/λ across three primes, with grokking at a near-constant memorisation norm. A basis-faithful FSD generalisation transfers to the non-abelian group S₅.

Pre-registered study of adaptive-depth (early-exit) latent world models across nine DeepMind Control tasks. Introduces the shallow penalty ρ = err(shallowest-exit rollout) / err(full-depth rollout) under matched K=1 and K=4 training. Depth helps rollouts on 6/9 tasks (ρ up to 4.7×), but 2/9 invert: shallow beats the full stack. An ablation shows the inversion is created by early-exit supervision, not the dynamics: the routability catch-22. ρ's sign predicts whether CEM planning benefits from depth.

A statistics-driven routing layer that selects among 6 function bases (Fourier, Wavelet, Chebyshev, RBF, Rational, Sigmoid) using distributional input statistics (mean, variance, skewness, kurtosis, entropy, sparsity) with zero routing supervision. Emergent specialization measured via per-layer Shannon entropy dynamics across training.

Pip-installable toolkit for routing entropy, expert utilisation, and collapse across OLMoE, JetMoE, and Qwen1.5-MoE. Includes domain-matched prompts (math/code/general), bootstrap confidence intervals, L_div LoRA training, MMLU evaluation, and a Gradio demo. Key finding: collapse is architecture-dependent, not universal.

Research assistant in Prof. Anna Choromanska's lab (NYU) working with postdoc Haoran Zhu on AD-LiST-JEPA, a Joint Embedding Predictive Architecture for automotive LiDAR object detection and occupancy completion/forecasting on the Waymo Open Dataset. Completed SSL pretraining pipeline; running downstream OCF finetuning experiments.

Evolutionary REINFORCE meta-learner over discrete MoE topology space (1-3 tiers × 2-5 experts/tier), eliminating hand-designed architectures. Discovered topology (3,3,2) achieving 100% task success at 11.5M params vs. 23.3M for fixed Hierarchical MoE: 2× parameter efficiency at equal performance.