Research

ENABOL: On-Device ML Training and Backpropagation

Jun. 2024 – Present · Related paper: ENABOL, DATE 2027 (under review)

• Developed an end-to-end Python/C++ HLS flow extending hls4ml, enabling on-device backpropagation and fixed-point weight updates directly on resource-constrained edge hardware.
• Formulated a Lipschitz-bounded operator projection method in PyTorch, mitigating gradient explosion and maintaining training convergence under strict 8-bit/16-bit fixed-point precision.
• Built data preparation pipelines to generate synthetic calibration distributions, achieving training parity with FP32 baselines while reducing hardware memory footprint by over 40%.

NetSuRF: Fault-Tolerant Toolchain for Edge-AI Accelerators

Jun. 2024 – May 2025 · Related paper: NetSuRF, IEEE TCAD 2026 (under review)

• Architected an automated Python framework applying selective redundancy (TMR) across neural network layers to harden edge models against bit-level transient faults under tight PPA budgets.
• Constructed a statistical weight-sensitivity profiler evaluating fault propagation across model parameters to minimize protection overhead.
• Synthesized multi-dimensional trade-off curves across power, performance, and area (PPA), validating model resilience across edge FPGA and ASIC deployment targets.

Real-Time Multimedia Processing Datapath and UVM Verification

Feb. 2024 – Mar. 2024

• Implemented a pipelined Sobel edge-detection engine with grayscale conversion and an asynchronous FIFO-backed UDP packet parser for real-time PCAP multimedia streaming.
• Built a modular UVM verification environment from scratch with constrained-random stimulus and functional coverage monitors, achieving 100% functional and code coverage.