How do we make large-scale personalization both efficient and reliable?
I work on LLM post-training and generative recommendation, with a growing
focus on making LLM inference fast and cheap enough to run at real production scale.
At Comcast, this means building knowledge distillation and post-training pipelines for user profile generation,
using LLMs to solve cold-start recommendation, and serving all of it efficiently enough to run over 5 to 10M users.
Generative & sequential recommendation
LLM-generated synthetic interactions and LLM-guided negative selection for cold-start and long-tail items (CIKM 2026, EACL 2027, NAACL 2027).
LLM post-training at scale
SFT, DPO, and distillation to train small models from larger teachers, cheap enough to deploy (NAACL 2027).
Efficient inference
Self-hosted vLLM serving that runs LLM personalization at roughly 650× lower cost than managed APIs.
News
Nov 2026Presented “From Head to Tail: LLM-Augmented Sequential Recommendation and Trade-off Optimization” at GRAIL @ CIKM 2026. [paper]
Oct 2026Two papers submitted to the NAACL 2027 Industry Track: “LLM-Driven Segment Discovery” [pdf] and “Candidate Grounding Matters: LLM-Guided Negative Selection for Production Recommenders” [pdf]. Under review
2026Submitted “From Head to Tail: LLM-Augmented Sequential Recommendation and Trade-off Optimization” to EACL 2027. [pdf]Under review
2026“Evaluating ASR Quality at Scale on TV Entertainment Platforms” accepted to GEM @ ACL 2026. [paper]
20252 patents filed at Comcast, on content filtering and on content discoverability for TV systems.
Sep 2025Presented “Simulating Discoverability for Upcoming Content on Entertainment Platforms” at RecSys 2025 in Prague. [paper]
2025“Content Moderation in TV Search: Balancing Compliance, Relevance, and User Experience” accepted to SIGIR 2025. [paper]
Selected Papers
UNDER REVIEW
From Head to Tail: LLM-Augmented Sequential Recommendation and Trade-off Optimization
Adeep Hande, Shaghayegh Agah, Neeraj Sharma, Yejin Kim, Karina H. Halevy, Sardar Hamidian
Recommendation and Arbitration team · Washington, D.C.
LLM post-training for user profile generation: built a teacher-student post-training pipeline for a 3B-parameter student model, using SFT with auxiliary reasoning objectives, DPO for alignment, and off- and on-policy evaluation. The fine-tuned models then act as teachers for downstream recommendation tasks.
LLM data augmentation for cold-start recommendation: designed an LLM-based pipeline injecting synthetic user histories for cold-start items, achieving an 80% relative NDCG improvement while maintaining overall quality (CIKM 2026, EACL 2027).
Efficient inference at scale: built a self-hosted vLLM serving setup on 4×A100 GPUs that reaches about 326k tokens/sec for user profile inference across 5 to 10 million users, roughly 650× cheaper than managed APIs at that scale.
Xfinity Campaign Manager forecasting: built campaign-level time-series forecasting models and ETL pipelines (Spark, Databricks, SQL) across EU and NA markets, with automated dashboarding for weekly media-buy pacing and budget allocation.
Content discoverability simulator: architected a large-scale simulation using LLM-generated synthetic queries and historical clustering to evaluate upcoming content in the Xfinity TV ecosystem (RecSys 2025).
Machine Learning Graduate Research Intern, Comcast
May 2023 – Dec 2023
Washington, D.C.
Xfinity ASR evaluation tool: built an end-to-end speech evaluation framework benchmarking ASR quality (WER, embedding-based metrics) across diverse user demographics and acoustic conditions, scaling annotation throughput from 1K/day to 100K/week via automated proxy metrics.
Search quality evaluation: deployed a hybrid system combining heuristics and LLMs for real-time moderation of 5M+ daily voice queries; introduced a time-aware dynamic lexicon reducing false positives by 35% (SIGIR 2025).
Education
Indiana University Bloomington
M.S. in Data Science
May 2024
Indian Institute of Information Technology, Tiruchirappalli
B.S. in Computer Science
Apr 2022
Skills
ML & NLPLarge language models, LLM fine-tuning (SFT, DPO, RLHF), knowledge distillation, recommendation systems, ranking, RAG, information retrieval, content moderation, bias & fairness, ASR evaluation
100+ peer reviews across academic venues including InterSpeech, SIGIR, COLING, ACL ARR, and ICASSP.
Outside Research
Outside of work, I follow Formula 1 and travel when I can. Recent trips include Peru, Mexico, Guatemala, Spain, Czechia, Austria, Germany, and Thailand.
After losing about 100 lbs, I got into hiking, running, and powerlifting:
Hiking: summited Volcán de Fuego in Guatemala, and my first 5,000 m peak, Pallay Punchu, in Peru.