Developed a JIT NPU runtime that overlaps CPU graph compilation with NPU inference and grows the KV cache with context: 91.98% less foreground compilation time and 33.05% lower average KV memory, while preserving model quality and decoding. Cached sorted compiler dependency vectors to reduce CPU sorting time by 94.2%, and built CPU/GPU/NPU benchmarks to guide device placement.

Measured how background filesystem reads interfere with a fixed LLM inference workload: median latency increased 17.1%. Windows PerfMon showed stable working sets and page-fault counts plateauing near 29.5K, supporting shared-resource contention as the explanation rather than working-set thrashing.

Built an LLM-driven pipeline over 50K+ Linux kernel commits to generate Clang-Tidy analyzers. Clustering and compiler-feedback repair produced 4× more valid checkers at 73% precision. A full-kernel validation harness uncovered 43 long-latent bugs with an average age of 4.7 years.

Contributed to experimental validation and performance analysis of CASH, which predicts workload–hardware affinity for CPU/GPU placement. In 5,000-task simulations using MIT Supercloud traces, the scheduler achieved 97.1% resource-matching accuracy, 46.5% lower response time, and 36.8% lower modeled energy than Meta-RHDC.

Coauthored a study using BEIT-style masked image pretraining to learn from unlabeled laryngoscopic images before fine-tuning for disease classification. The paper reports 93.33% accuracy on a three-class task using a ViT-Tiny backbone with 5.06 million parameters.

Coauthored a study of a voltage-tunable near-infrared filter combining polymer-dispersed liquid crystals with a Fabry–Perot cavity. The paper reports a tuning range of 913–969 nm and 99% transmittance at 927 nm, with a 24 nm full width at half maximum.