PROJECT INDEX

Research & industry

Questions, design decisions, and contributions across research and engineering.

TaRo

Purdue UniversityCVPR 2027 · Under review

Temporal candidate reuse for streaming video-language models

Reuse promising KV candidates across video frames, then rescore them with the current query while retaining the complete KV history. TaRo achieved 1.37–3.58× LLM-prefill speedups over ReKV in the evaluated settings.

TaRo Figure 3: video frames pass through a vision encoder into TaRo-equipped LLM layers. Stored indices 1, 5, 11 and current causal keys 14, 15 are rescored with the current query. Top-k feeds attention; a full-history search refreshes the bank when the query changes or the bank reaches its age limit.
Original paper figureClick to enlarge

EdgeVQA

Purdue · POSTECH collaborationDAC 2027 · Under review

Sustaining video and answer rates on bandwidth-limited devices

Process video frames and verify draft answers in one model-weight sweep, with adaptive chunk sizing to keep answers moving. The evaluated design reduced modeled bandwidth for 10 frames/s and 3 tokens/s from 55.5 to 23.0 GB/s.

An adaptive scheduler combines incoming-video processing and draft-answer verification in a shared weight sweep. Answer progress guides the next video chunk size.
Concept diagramClick to enlarge

HEED

Purdue UniversityResearch project

Workload prediction and DVFS for energy-efficient edge AI

Predict the required network depth from temporal continuity, then adjust voltage and frequency to meet each deadline. HEED saves 16.17–17.55% energy versus EENet at matched accuracy in the evaluated MCU audio streams.

HEED Figure 1: overlapping audio segments pass through feature extraction and pre-classifier prediction, then a classifier with early exits and an in-classifier fallback; a task timeline marks voltage/frequency scaling points.
Original paper figureClick to enlarge

TESSERA

Purdue UniversityIISWC 2026 · Published

Workload-driven heterogeneous NPU exploration

Explore how large, small, and special-function tiles can work together in heterogeneous NPUs. I built the preliminary exploration prototype and co-authored the IISWC 2026 paper.

TESSERA Figure 4: workload and architecture inputs plus calibration feed a cost-aware compiler, a tile and chip-level simulator, and a design-space-exploration engine.
Original paper figureClick to enlarge

Hardware-aware compression

LG Display · 2023–2024RTL / ASIC

Vector quantization and memory dataflow co-design

Co-design vector quantization and memory dataflow for per-frame display compensation. Achieved 8:1 data reduction and reduced DDR channels from two to one.

View project flow
Project diagram

From compensation blocks to compact indices

Vector quantization example: the input block 8, 7, 8, 8 matches codebook row 1, which stores 8, 7, 8, 7. The block is stored as index 1; a DDR index retrieves the shared SRAM vector to reconstruct the block.

Each block of compensation values is represented by the index of its nearest shared code vector. During operation, indices in DDR select vectors from the SRAM codebook to reconstruct the values. Green traces one example block; numbers and block sizes are illustrative.

Vector quantization and runtime memory placement

Real-time thermal estimation

LG Display · 2025RTL handoff

From a physical model to a hardware datapath

Turn video-derived current maps into pixel-level temperature estimates using 2D heat diffusion. Adapted the model for real-time hardware and completed RTL handoff.

OLED degradation modeling

LG Display · 2024–2025Applied ML

Predictive models and interpretable compensation parameters

Model OLED degradation with machine learning and generate compensation parameter maps. Explainable AI helps identify the factors behind the predictions.