Theme
Memory Wall in Edge Inference
I am interested in how transformer-based perception modules suffer from memory bandwidth and locality bottlenecks on edge devices such as Jetson-class platforms.
Research
Research Direction
My current research direction centers on efficient AI inference, especially how model-level choices interact with compiler passes, runtime policies, and hardware realities.
Hardware-aware co-optimization of quantization and runtime schedulingEfficient inference for embodied AI and vision agents on edge devicesBridging the semantic gap between model abstractions and low-level execution
Research Overview
This page captures the current technical direction behind the site: efficient inference, system optimization, and hardware-aware execution.
Theme
Quantization and Operator Fusion
Rather than treating quantization and fusion as isolated stages, I want to study them as a shared search space shaped by instruction-level and hardware-level constraints.
Theme
Runtime Scheduling
I care about how runtime policies, especially CPU-GPU coordination and dynamic scheduling, affect deterministic inference behavior in resource-constrained systems.
Agenda
Current Agenda
- Hardware-aware co-optimization of quantization and runtime scheduling
- Efficient inference for embodied AI and vision agents on edge devices
- Bridging the semantic gap between model abstractions and low-level execution
Outputs
Outputs and Materials
- One manuscript in preparation on multimodal video captioning
- Research proposal on hardware-aware AI systems optimization
- Project portfolio spanning CUDA, RAG systems, and multimodal inference