From frontend
to silicon.

01

APPLICATION

Perception that keeps a car safe.

Camera + LiDAR + thermal fusion — so a vehicle sees what a camera alone can't, on hardware that fits in the car.

83.9%pedestrian detection
up from 40.4%

2.7fpsreal-time on
Jetson Nano

Thermal-RGB fusionLiDAR + 2DLane detectionGTA5 digital twin

PerCom · IEEE Access · Sensors HyundaiTesla

02

MODEL & SOFTWARE

Learning that costs almost nothing.

Spiking neural networks meet self-supervised learning — plus the compiler that turns a model into tuned hardware code, automatically.

1stcomprehensive SNN-SSL
hardware framework

FPGA+ASICauto-generated
from one model

Spiking NNSelf-supervisedPruningQuantizationTVM codegen

DAC · IEMEK · MWSCAS NVIDIAETRI

03

ARCHITECTURE

Cores & accelerators, designed — not just used.

I design the datapath and the control: a spiking NPU, a RISC-V clock controller, a custom Cortex-M0+ core with my own µSIMD.

4×MAC throughput
via custom µSIMD

IEEETVLSIS³A-NPU
journal paper

Spiking NPURISC-V controlCustom ISA / µSIMDSparsity zero-skip

TVLSI · COOLChips · custom ISA ArmAMDNVIDIA

04

MEMORY & CIM

Win at the memory wall.

Moving data costs the most energy — so I attack it directly: on-chip memory that adapts in real time, and compute done inside the memory with RRAM.

RRAMcompute-in-memory
fabricated (MPW)

runtimeadaptive on-chip
memory allocation

Compute-in-memoryAdaptive on-chip memBuffer minimization

RRAM CIM tape-out · IEEE Access · ICCE SamsungSK hynix

05

SILICON

It's not real until it's fabricated.

Every idea above ends here — a taped-out chip, a hand-routed PCB, a physical device. RTL to GDS.

4chip
tape-outs

3process nodes
28nm · 130nm · MPW

RTL → GDSSamsung 28nm ×2SK 130nmCustom PCB3D printing

Full ASIC flow · home lab Every silicon team