Experience | Pengcheng Wang

Selected Research and Professional Experience

EmbodyX · Sep 2025 - Present · West Lafayette, IN

Investigating systems-level optimization techniques for accelerating LLM and VLM training and inference in foundation model development.
Exploring compute-efficient adaptation and parameter-efficient fine-tuning methods for deployment on constrained platforms.
Designing system-level solutions for robust inference performance in real-world deployment environments.

Purdue University · Jan 2025 - Present

Characterized the VLM inference design space across token, width, and depth dimensions to identify efficiency-accuracy trade-offs.
Developed compute-adaptive VLM techniques that meet computational budgets while preserving accuracy under degraded inputs.
Designed learning-based inference schedulers that jointly consider content, noise, latency, and per-sample resource allocation.

Sunlune · Jan 2025 - Aug 2025 · Santa Clara, CA

Developed and validated kernel, runtime, and driver frameworks for domain-specific AI accelerators.
Integrated LLM inference kernels and optimized runtime scheduling workflows for Llama-family models on custom AI hardware.
Conducted performance profiling and cross-platform debugging to identify bottlenecks and improve kernel execution efficiency.

Sunlune · May 2024 - Jan 2025 · Santa Clara, CA

Developed AI-driven design automation flows for high-performance digital ASIC design.
Designed deep reinforcement learning models for logic synthesis and technology mapping, improving delay-power trade-offs across circuit designs.
Integrated reinforcement learning with human feedback to encode IC design expertise into optimization models.

Purdue University · Aug 2022 - Jan 2025

Proposed an adaptive 3D object detection system for LiDAR point clouds that meets runtime latency targets while maximizing detection accuracy on NVIDIA embedded GPUs.
Analyzed bottlenecks in 3D detection pipelines, including voxelization, voxel feature encoding, and 3D spatial feature extraction.
Introduced configurable control knobs across the 3D encoder, CNN backbone, and detection head for fine-grained accuracy-latency trade-offs.
Evaluated on NVIDIA Jetson AGX Orin and Xavier, achieving higher mAP than strong baselines while sustaining real-time throughput.

Teaching Assistant, Purdue University, Department of Agricultural and Biological Engineering, Jan 2024 - May 2025
ABE 591: Machine Learning for IoT and Computer Systems, Spring 2024 and Spring 2025.
Teaching Assistant, Tongji University, Department of Electronic Science and Technology, Sep 2014 - Jan 2017
Semiconductor Physics, Fall 2014, Fall 2015, and Fall 2016; Electronics and Digital Technology, Spring 2015; Electromagnetic Fields and Waves, Spring 2016.