01 Work

Selected projects

Case studies with the measurement conditions attached. Where a number is missing, the work has not earned it yet.

4 projects · 2 with published numbers

active 2026

COLIDE

Custom CUDA kernels for a CNN-BiLSTM intrusion-detection model, with an on-device LLM that explains alerts without ever sitting in the detection path. A systems study — and an exercise in not overclaiming.

CUDA
HPC
IoT Security
LLM
Systems

Case study Code

Table 1 At a glance results

BoT-IoT macro-F1: 0.9780 ± 0.0033; Sealed multi-seed test, n=5 (seeds 42–46)
Blocks 1/2/4 vs matched PyTorch: 3.24×–6.55×; Operator-for-operator, RTX 3050
Block 3 kernel progression: 7.55×–9.50×; Naive → FP16 half2 gate packing, five sessions
Alert dispatch: 16.60 µs p99; Dispatch only — generation runs off the detection path

Active

2026

Slipstream

An LLM inference engine built from scratch, where the research sits in the scheduler — an output-length-predictive policy that decides which sequences get to occupy a full KV cache.

LLM
Inference
Scheduling
Systems
Triton

Read the case study

Active

2026

qgemm-mx

Recovering block-scaled FP4 throughput on GPUs that have no native FP4 multiply. The memory saving already ships; the speed is being left on the floor, and this measures exactly where it goes.

CUDA
Quantization
Hopper
Kernels

Read the case study

Active

2026

RISC-V parameter extraction

Auditing whether a published extraction benchmark measures what everyone reads it as measuring. It did not — the correction was filed upstream, and the fixes it produced were merged.

RISC-V
Measurement
Open Source
Evaluation

Read the case study