Software · agents · accelerated computing

Build the runtime, then prove what it does.

I am interested in software that exposes its own evidence: coding agents with replayable traces, GPU kernels with roofline measurements, game engines with exact correctness oracles, and services with visible objectives and error budgets.

Hardware-facing path

Kernel behavior to accelerator design

The useful portfolio bridge is a quantitative chain from algorithm to dataflow, memory traffic, RTL, and system cost.

01 · Kernel

Measure the operation

Arithmetic intensity, occupancy, divergence, locality, and numerical error.

02 · Dataflow

Choose what stays put

Weight-, output-, or row-stationary schedules and explicit SRAM/HBM traffic.

03 · RTL

Build the block

Quantized MAC array, scratchpads, DMA/AXI, control, cycle model, and assertions.

04 · System

Prove the tradeoff

Timing, area, power proxy, utilization, compiler schedule, throughput, latency, and cost.