Research

arXiv preprint (cs.IR)

Carbon-aware product recommendation when PCF labels are sparse: a retrieval-augmented pipeline transfers supervision from the Carbon Catalogue to a large catalog via similarity search, few-shot LLM prompting, and nearest-neighbour fallback, then post-hoc re-ranks BPR, NeuMF, and LightGCN scores with a single tunable λ trading engagement for estimated footprint. Evaluated on Amazon Reviews (Home & Kitchen, Sports & Outdoors, Electronics); λ sweeps yield Pareto frontiers showing substantial carbon reductions at small engagement cost, with headroom varying by model and category. With Anders Vestrum and Jorgen Bergh.

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 that samples diverse candidates, builds an NLI-based semantic graph, and gates on a global uncertainty score before optionally asking one binary clarification and regenerating. We introduce AmbigSum (500 CNN/DailyMail-derived examples with clarification annotations) for evaluation. On the 350-example test split, the full pipeline achieves mean document entailment 0.567 (ask rate 0.42), improving over a greedy BART baseline (0.498; Δ = +0.069, p < 0.001), with larger gains where the gate triggers clarification.

UC Berkeley (INFO 288)

Tests whether the Weiss et al. global healthcare travel-time map tracks DHS-reported access barriers equally across Zambian provinces and seasons, using two independent DHS waves. Performance differs sharply by province; we train a route-level model that augments travel time with precipitation, route geometry, and accessibility features to predict cluster-level shares reporting distance as a major barrier, exposed as seasonal risk scores in an interactive dashboard.

NTNU x NINA

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