Research

UC Berkeley (EECS 282)

Trained GPT-style Transformer models to study robustness of in-context learning under noisy demonstrations across synthetic regression and NLP benchmarks. Evaluated curriculum-based noise schedules on GPT-Neo 2.7B and Llama-2 7B.

UC Berkeley (CS 288)

Clarification-aware abstractive summarization system that uses multi-sample disagreement and NLI-based uncertainty estimation to decide when to ask a targeted clarification question. The pipeline samples diverse candidates from BART, builds a semantic entailment graph, gates on global uncertainty, localizes the riskiest sentence, and regenerates conditioned on the resolved answer. Full pipeline improves over greedy baseline by +0.079 mean entailment (p = 0.0001) on a 40-example test set, with ablations confirming that both the selection policy and targeted localization contribute.

UC Berkeley (EECS 294)

Carbon-aware re-ranking framework for e-commerce recommendations. Estimates product carbon footprints by combining retrieval from the Carbon Catalogue with few-shot LLM prompting (Spearman 0.853), then applies a tunable post-hoc re-ranking policy over BPR, LightGCN, and NeuMF candidates (RecBole). On Electronics, BPR sustains full NDCG@10 through λ = 0.90 while cutting average footprint by over 35%, revealing a plateau-then-cliff trade-off structure across all models.

UC Berkeley (INFO 288)

Modeling barriers to healthcare access using DHS survey data (16k+ responses) combined with geospatial travel-time, rainfall, and infrastructure data. Achieved Spearman ~0.53 and identified key drivers of access inequality in rural regions.

NTNU x NINA

Bachelor thesis project developing a scalable system to manage and visualize passive acoustic environmental data, improving automated biodiversity monitoring workflows.