Evaluation of agentic AI systems generation quality, agentic behaviors, and safety.
Senior Machine Learning Researcher, Apple · Seattle, WA
My research addresses a central problem in modern AI: how to rigorously evaluate agentic systems once they move beyond single-turn answers into long-horizon, multi-step tool-use with error recovery. In this regime the conventional axes of quality — correctness, completeness, relevancy — are necessary but no longer sufficient. I develop compound measures of agent behavior, including user utility, user-perceived defects, and long-term rollouts. I also pursue a complementary line on cost and latency-aware routing that reserves frontier-model capacity for the hardest tasks while holding generation quality invariant. Together these threads aim to make agentic systems measurably more trustworthy, efficient, and accountable at production scale.
I pursue this work as a Senior Machine Learning Researcher in Apple's Human-Centered AI organization, where I scale agentic evaluation for Apple Media Products. Earlier, as an Applied Scientist at Amazon, I built LLM-driven conversational systems, large-scale insight generation, and multi-modal fraud detection across the Returns & Recommerce and Seller Partner Services organizations. I hold an M.S. in Computer Science from the University of Illinois at Chicago, where I was advised by Xinhua Zhang and worked at the intersection of few-shot machine learning and computer vision.
(Coming Soon)
Citation counts via Google Scholar, as of July 2026.
Input
Depth
A survey of supervised, weakly-supervised, and unsupervised approaches to predicting dense depth from a single RGB image, comparing methods and outlining open directions.
@article{bhoi2019monocular,
title = {Monocular Depth Estimation: A Survey},
author = {Bhoi, Amlaan},
journal = {arXiv preprint arXiv:1901.09402},
year = {2019}
}
Frame
Action
A survey of action localization in video — determining what action is performed, and when and where — across algorithms, datasets, and the most promising directions.
@article{bhoi2019spatiotemporal,
title = {Spatio-temporal Action Recognition: A Survey},
author = {Bhoi, Amlaan},
journal = {arXiv preprint arXiv:1901.09403},
year = {2019}
}
An exploration of pre-processing, features, and memory-network methods for classifying sentiment toward a specific aspect within a sentence.
@article{bhoi2018various,
title = {Various Approaches to Aspect-based Sentiment Analysis},
author = {Bhoi, Amlaan and Joshi, Sandeep},
journal = {arXiv preprint arXiv:1805.01984},
year = {2018}
}
Content
Stylized
A comparison of neural and universal style-transfer approaches, focused on real-time transfer and generalization to unseen styles.
@article{majumdar2018comprehensive,
title = {A Comprehensive Comparison between Neural Style Transfer
and Universal Style Transfer},
author = {Majumdar, Somshubra and Bhoi, Amlaan and Jagadeesan, Ganesh},
journal = {arXiv preprint arXiv:1806.00868},
year = {2018}
}
Peer reviewer across ICML, EMNLP, NeurIPS, KDD & COLM 2026 venues.
Scaling foundational, trustworthy agentic evaluation systems across Apple Media generative AI features.
LLMs for conversational returns, large-scale insight generation, and multi-modal fraud detection; evaluation of generative-AI listings and Seller Assistant.
Image classification and segmentation for automotive damage assessment.
Machine learning & computer vision; advised by Xinhua Zhang. Thesis: Invariant Kernels for Few-shot Learning (Outstanding Thesis Award). Introduced invariant representations to few-shot learning via image tiling, shift-invariant kernels (Gaussian/RBF), Nyström's method for low-rank approximation. Proven with strong theoretical guarantees.