Peer-reviewed publications in ML for materials science, plus in-depth reviews of systems-for-ML research — distributed training, ML infrastructure, inference serving, and emerging AI systems, from Ray and TensorFlow to vLLM, MegaScale, and DistServe.
ML framework for predicting tensile properties (yield strength, elongation) of miniaturized SS-316 nuclear specimens from sparse, imbalanced data. GAN + SMOGN data augmentation; Random Forest / XGBoost reached Pearson r > 0.98, establishing guidance for matching augmentation strategy to target property. Co-author, with INL collaborators.
Co-authored work demonstrating a transfer-learning deep-learning framework that outperformed four state-of-the-art models for cross-scale defect segmentation in nuclear materials.
Real-time exercise-monitoring approach using BlazePose for pose estimation, yoga-pose classification, and automated activity reporting.
Methods for detecting blood chambers in echocardiogram video frames using a U-Net segmentation approach.
The models below run in live broadcast pipelines. My work on them is the same output, faster: profiling, FP16 precision, and GPU-pipeline optimization that lifts throughput while the predictions stay identical. Methods stay under the hood — the numbers are the point.
real-time tracking · basketball dunk
Optimized the sequential tracking pipeline for the live dunk-show metrics — throughput rose from 45 to 61 FPS with accuracy held, so metrics land in real time for the show's team.
fine-tuned tracker · skateboarding
Fine-tuned and profiled the autonomous skater tracker to reach up to ~79 FPS on an RTX 6000 with exact output parity — same masks, same decisions, just fast enough for live X Games judging.
ball tracking · pickleball line calling
Fine-tuned TrackNetV5 for broadcast ball tracking — F1 climbed from 95.1% to 99.1% — and served it under TensorRT at 120 FPS live, keeping ball detection reliable enough to stand behind on-air in/out calls.
fine-tuned segmentation
Fine-tuned for broadcast segmentation and moved to FP16 precision — smaller, faster inference with equivalent output quality.
video object tracking
Ran the SAM 3 video tracker at 60 FPS across a 2× RTX 6000 setup — a 3× throughput gain — keeping temporal tracking real-time on broadcast footage.
causal · snowboarding trick delimiting
Built a custom causal temporal model for snowboarding — it consumes only past frames (no lookahead, so it is real-time safe) and outputs a per-frame probability the rider is airborne, cleanly marking trick start and end. Optimized from 45 to 68 FPS.
pose estimation
Pose estimation inside the real-time tracking stacks, supplying keypoints that drive trick and metric computation.
transformer pose estimation
Transformer-based pose estimation in the Triton-hosted broadcast model zoo, feeding athlete keypoints to the trick- and metric-computation stacks alongside the detection and segmentation models.
semantic segmentation
Fine-tuned semantic segmentation that holds up in cluttered broadcast scenes — props, crowds, and occlusions in the dunk and skate pipelines.
human-centric vision
Fine-tuned human-centric vision for athlete body understanding, feeding the pose and metric layers of the sports pipelines.
detection & segmentation
The detection-and-segmentation backbone of the live pickleball line-calling stack — ball tracking and bounce localization behind the near-perfect on-air in/out calls.
Whisper · SAM Audio · Qwen TTS · VibeVoice
Speech and audio models reading venue sound at live events to gauge audience pulse — crowd reaction and excitement signals — from the arena's own soundtrack. The same stack drives a motocross rebroadcast-dubbing pipeline: ASR + speaker diarization, staged LLM translation (~99% accurate), and per-commentator TTS voices timing-fit to the original broadcast windows.
applied-ML infrastructure · multi-GPU
Orchestrates multimodal models — Qwen-3VL, NVIDIA Cosmos, SAM and SAM Audio, plus VLMs that detect and segment action sequences into ground-truth clip datasets — across multi-GPU systems with chunked, round-robin inference scheduling. Auto-labeling cut manual annotation time 90%+, turning raw broadcast footage into training-ready signals for every model above.
In-depth reviews of research on distributed training, ML infrastructure, inference serving, and emerging AI systems. Each card links to the full review PDF.
Ray's unified interface for task-parallel and actor-based computation with a dynamic execution engine.
Zero-shot video search using VLMs with prompt expansion and diversity-aware selection.
Flexible dataflow-graph architecture enabling scalable ML across CPUs, GPUs, and TPUs.
Google's TPU delivering 15–30× speedups for neural-network inference via domain-specific design.
Framework unifying ML input pipelines with systematic optimization across PyTorch, TensorFlow, and JAX.
Policy-driven retraining for continuously growing datasets with drift-based triggers and sample selection.
Compiler automating parallelism strategies for large-model training via ILP and dynamic programming.
Combines intra- and inter-batch parallelism with 1F1B scheduling and weight stashing.
Extends ZeRO-3 with NVMe offloading and memory-centric tiling to train trillion-parameter models.
Fault-tolerant training using adaptive pipelining and decoupled backprop to sustain throughput on GPU failures.
Meta's system scheduling ML across geo-distributed datacenters, achieving 98% GPU allocation efficiency.
Declarative inference API with automatic variant selection, ILP cost optimization, and vertical autoscaling.
Shows model size and training tokens should scale equally for compute-optimal training.
Iteration-level scheduling and selective batching, achieving 36.9× higher throughput than FasterTransformer.
PagedAttention uses OS-like paging for KV-cache management, achieving 2–4× throughput and 96%+ utilization.
Separates prefill and decoding onto different GPUs with bandwidth-aware placement, up to 7.4× higher request rates.
Hierarchical search with clustered indices for trillion-token RAG, achieving 9.33× speedup.
LLMs combined with memory streams, reflection, and planning for believable, emergent agent behavior.
ByteDance's system achieving 55.2% MFU on 12,288 GPUs via algorithm-system co-design.
Unified paging and heterogeneous batching with custom CUDA kernels, 4× throughput over vLLM.
Privacy vulnerabilities in ML components — filters, tokenizers, memorization detectors — beyond model privacy.
Holistic analysis of AI's carbon footprint — embodied and operational — across the ML lifecycle.