Applied AI Engineer

Implemented a training pipeline for correction-aware chain-of-thought reasoning models using supervised fine-tuning, rejection-based sampling, and Group Relative Policy Optimization (GRPO).

Integrated vLLM for inference-heavy stages and managed dynamic KV-cache flushing after backpropagation to improve training throughput. Ran exploration-strategy ablations with a validator and optimized the GRPO pipeline to achieve an approximately 20x speedup over the baseline Hugging Face Transformers forward loop.

Investigated uncertainty estimation for geological images using an ensemble of U-Net models. Used ensemble outputs to generate pseudo-labels for student-model training; dataset details are confidential.

Analyzed uncertainty-evaluation results, identified failure modes, formulated hypotheses about their causes, and designed experiments to test them. Reduced the area under the risk-coverage curve by 60% relative to the baseline on the evaluation dataset.

Researching optimism in agent AI security under adaptive defense, using evaluations where attackers and defenses adapt to one another.

Built a research pipeline using Evolution Strategies, perturbation, batched generation, sandboxed rollouts, and dual-GPU execution. Preliminary results indicate optimistic security conclusions, with ongoing experiments measuring how optimism changes with defense budget.

Compared a single-model baseline for modular addition with an ensemble of four small transformers using KL-softmax consistency.

Across eight controlled experiments, identified learning rate as the main factor driving generalization improvement. The ensemble achieved 100% accuracy compared with 12% accuracy for the single-model baseline.

Currently pursuing a Master's degree in Artificial Intelligence and Data Engineering. Expected graduation: 2027.

Graduated with a GPA of 4.4/5.0. Bachelor's thesis: Convolutional Kolmogorov-Arnold Network for Self-Supervised Visual Representation Learning. Benchmarked CKANs against convolutional baselines for SimCLR-style self-supervised learning and found comparable downstream performance at substantially higher compute cost.