Current Role: Tech Lead, DIGI @ University of Washington
Self-contained technical walkthroughs, mathematical derivations, and executable notebooks hosted at tutorials.pmanko.dev.
Implementation notes and profiling benchmarks for small open-weight models: LoRA supervised fine-tuning (SFT), Group Relative Policy Optimization (GRPO), on-policy distillation, and Quantization-Aware Training (QAT) with hardware roofline analysis.
Walkthroughs on training bidirectional cross-encoders for evidence admission and contradiction detection, covering temperature scaling, isotonic regression, Expected Calibration Error (ECE), and conformal risk control.
Reference implementations for multi-seed Wilson score intervals, Prediction-Powered Inference (PPI) confidence interval rectification, ceiling-normalized Cohen’s κ, and perturbation resilience scoring.