ML systems · Research · Open source
ML systems engineering. Compiler contributions. Applied AI research.
I work where AI models meet the systems that run them. My contributions have landed in PyTorch, OpenXLA, Apple's CoreMLTools, and Apache Airflow, spanning compiler optimizations, image-model conversion, and scheduler reliability.
At the University of Maryland's Child Development Lab, I develop ML and NLP pipelines for parent–child interaction research. My research also spans language-model reasoning and persona-agent evaluation, alongside academic service and engineering leadership.
Selected contributions
Concrete changes to shared tools: how models compile, which operations can convert, and how workers shut down reliably.
Extended an existing graph rewrite to eligible batched matrix operations, making the bias addition available to normal pointwise fusion. The patch preserves scaling and precision guards and adds generated-code regression checks.
Added conversions for deformable convolution, fixed-sampling ROIAlign, and bicubic resizing, including antialias variants. These contributions express source operations through existing Core ML building blocks and add numerical tests.
Fixed asymmetric copy elision in conditional code: equivalent branch arrangements could trigger unnecessary buffer copies. The contribution adds regression tests for both branch orders and a conditional inside a loop.
Fixed a LocalExecutor deadlock where a parent could wait for workers that were blocked on a full results pipe. Draining results while waiting for workers allows shutdown to make progress.
Language-model reasoning, human interaction, and agent-based evaluation. Each item identifies its authorship role and publication milestone.
Accepted · workshop proceedings track
NeurIPS 2026 · NeurReps Workshop
Sole-author work on language-model reasoning and representational geometry, accepted to the workshop's Proceedings Track.
Accepted · workshop paper
COLM 2026 · Efficient Reasoning Workshop
First-author work on supervision mismatch in process-reward-model distillation. Accepted as a non-archival workshop paper.
Published · meeting abstract
Developmental Psychobiology · 67(S1), e70105
First-author meeting abstract connecting machine learning with the study of parent–child affective synchrony.
Preprint · co-author
Population-scale persona-agent evaluation
Co-author of research on evaluating AI systems and digital products with diverse simulated users.
Area Chair for NeurIPS 2026 AIDaR and EMNLP 2026 GroundLM workshops; Program Committee member for AAAI/ACM AIES 2026.
NAIRR Pilot review-panel appointment and State & Regional Judge for the White House AI Presidential Challenge. Service details are listed separately.
Engineering leadership, applied ML education, research mentorship, and community building at the University of Maryland.
I am an undergraduate at the University of Maryland, College Park, and a research assistant at the Child Development Lab. You can find my open-source work on GitHub as SakshamKapoor2911.
Undergraduate student
Research assistant working on ML/NLP pipelines for parent-child dynamics research
Research collaborator and co-author on persona-based simulation and AI evaluation
Vice President
Fellow in UMD's student-run incubator community