Independent researcher exploring epistemic uncertainty, belief formation, and policy change in deep learning systems — with a particular interest in whether a model's observed behavior actually tracks its intended objective over time, versus fitting surface patterns that look like it does.
- ERATO — ERATO — tracks behavioural and policy-level change across post-training checkpoints (SFT/DPO/GRPO): fixes, regressions, durability, and whether preference-margin shifts (RPMS) transfer to held-out prompts or reflect pair-specific fitting. Representation-level analysis is planned future work.
- bayzflow — converts PyTorch models into Pyro-backed Bayesian models for uncertainty-aware prediction.
By day I build dynamical-systems/manifold-trajectory analysis for operational data; the same instinct — trajectory over snapshot, and separating real signal from second-order noise.