Amartya Banerjee

Applied Scientist, Prime Video & Studios CoreTech

Representation Learning · LLM Post-Training · Information Retrieval · Recommender Systems

Investigated whether preference feedback can supervise retrieval without touching the underlying language model. Formulated a feedback-supervised, contrastively trained two-tower objective that embeds LLM-generated interest representations and customer viewing histories in a shared space. Established consistent full-catalog gains in recall, NDCG, and MRR over a frozen semantic-retrieval baseline, showing that feedback-driven personalization is attainable without LLM fine-tuning.

Applied Scientist, Prime Video & Studios CoreTech

Causal Inference · Reinforcement Learning · Decision-Making under Uncertainty

Studied demand forecasting for dynamic pricing when the training data is confounded by the historical pricing policy that generated it. Combined counterfactual demand modeling with reinforcement learning, and derived tunable online correction layers that debias logged-policy effects while adapting to distribution shift. Improved WAPE, MAPE, and MAE, and materially reduced how often models must be retrained.

Machine Learning Research Intern

Information Theory · Nonparametric Learning · Robustness

Worked on anomaly detection inside Howso's nonparametric, nearest-neighbor learning framework, scoring observations by surprisal rather than by scaled distance. Benchmarked it across a broad set of tabular datasets against a wide range of baselines, spanning traditional methods such as Isolation Forest through to deep models like Deep SVDD, and studied how well it holds up under adversarial perturbations. The method is now part of the open-source Howso Engine, and the work led to a first-author preprint.