Hitesh Golchha

I am currently in an Applied Scientist at Amazon. I have work and research experience in the fields of NLP, Conversational AI (language generation and understanding), Reinforcement Learning and Representation Learning from companies like Amazon, Flipkart and research labs at UMass Amherst, IIT Patna and Bar Ilan University.

I love making educational content on my YouTube Channel ExplainingML. Some videos include LLM papers (Deepseekv3, Reformer) or proofs of RL algorithms for LLM Post Training (GRPO/Dr GRPO, DPO, PPO, TRPO, Policy Gradients, Key concepts in RL) as a Playlist.

Email  /  Google Scholar  /  Twitter  /  Github  /  YouTube  / 

Language-Guided Exploration (LGE) injects language priors into sparse-reward RL by using a GUIDE model to rank and prune actions for an EXPLORER agent. GUIDE is trained via contrastive learning to prefer task-relevant actions, dramatically shrinking the effective decision space. On ScienceWorld, this yields faster learning and higher returns than vanilla RL, Behavior Cloning, and Text Decision Transformers.

We develop a system that transforms a generic customer care response into a 'Courteous' one. The model takes into context the emotions and content of the conversation history as well while generating and is trained on MLE+RL.

We compare multi-task approaches with standalone and pipeline ones for Spoken Language Understanding tasks on ATIS, TRAINS and FRAMES datasets.

We develop a linguistically informed, self-supervised model to identify customer-to-customer suggestions in Hotel and Electronics review datasets.

As an Applied Scientist at Amazon, I lead several ML projects for Amazon’s Cross-Border (Global Store) business, which enables international product discovery and purchasing across countries and marketplaces. My work spans Agentic systems, Deal-forecasting models, large-scale retrieval, Scalable Product Clustering deployed in production and directly supporting multi-million-dollar revenue streams.

Built LLM-powered QA, summarization, and ML review systems for 400+ Model Risk & Governance (MRGR) users using LLaMA-13B with advanced RAG pipelines (section-aware chunking, hybrid parent-child retrieval, reranking, map-reduce and refine summarization). Also evaluated Information Extraction, OCR, operations and Fraud Detection Models in Consumer Investment Banking for Firmwide risks.

Probing Product representations from vision, text, and multi-modal models to find Products having similar export-eligibility. Also made contributions to other ML projects within the team performing diverse subset sampling.

Multiple research tasks in the Conversational AI: Intent detection, slot filling, multi-intent sentence segmentation, data augmentation, ASR beam rescoring. Challenges included handling multilinguality, class imbalance, miscalibration, augmentation and robustness. The voice assistants we developed are used by millions of Indians for E-commerce shopping every day. I also participated in hackathons on unsupervised reverse image search and question answering.

– Text-based Games using RL : Using LLMs, RL and Contrastive Learning for playing text-based science games.
– Energy based models for Multilabel Classification : Exploring architectures for Energy Based Models for Multilabel Classification which can model contextual label dependencies.
Worked with researchers from IESL lab under Prof. McCallum and IBM Research.

Research in the domain of conversational agents: NLU and Style Transfer in Response Generation.
Advisor: Prof. Asif Ekbal and Prof. Pushpak Bhattacharyya.

Research in Coreference Resolution of Propositions. I got an exposure to several tasks in NLP associated with creation of Natural Language Knowledge Graphs.
Advisor: Prof. Ido Dagan.

Research in Suggestion Mining, Question Answering
Advisor: Prof. Asif Ekbal and Prof. Pushpak Bhattacharyya.