projects
Projects
School work and individual work, filterable by area. Case studies land here as they are written.
School work
Fast Robots: Sensor Fusion and Autonomous Navigation
RoboticsA hand-built differential-drive robot that localizes with a Bayes filter and pulls off high-speed maneuvers by Kalman-filtering sensor data faster than the sensors themselves update.
case study →
Huey: Autonomous 3 lb Combat Robot
RoboticsA beetleweight combat robot that finds and engages its opponent on its own — vision-based detection feeding an orientation controller, in an environment built to destroy it.
weight class: 3 lb
in progress
case study →
Parallel Sparse Matrix–Matrix Multiplication
Systems · HPCA parallel SpGEMM built on Gustavson's algorithm, reaching ~21× speedup on a 32-core Xeon system by attacking load imbalance and memory-bound accumulation.
speedup: ~21× on 32 cores
case study →
Super Gold Hunters
GraphicsA first-person maze game whose 3D renderer we wrote from scratch in Java — no engine and no libraries, just a ray cast down every one of 640 screen columns.
dependencies: none
case study →
Individual work
Bits → Messages
EmbeddedA serial-communication simulator whose engine also drives a real device node — the same UART runs the diagram in the browser and the /dev/ttys* you open with screen.
protocols: six, compared
case study →
Blaze and Heat
RoboticsA combat-robotics game with no hit points: the spinner stores real rotational energy, and the collision that dumps it is what decides whether armour shears off.
stored blade energy: 4302 J at 1910 RPM
in progress
case study →
Electrons → Instructions
EmbeddedAn interactive eleven-level descent from charge drifting in copper to a running program that puts a bit back on a pin — and every level can be broken on purpose.
levels: 11, each earned
case study →
Emulation-Based Fault Injection Harness
EmbeddedA campaign harness that glitches bare-metal firmware in emulation to measure which countermeasures actually stop a secure-boot bypass — across two instruction sets, with every result cross-checked against a second emulator.
exhaustive search: 95.9M fault pairs
in progress
case study →
Local-First Imitation-Learning Platform
RoboticsThe whole imitation-learning loop on one laptop — teleoperate a manipulation task in the browser, record demonstrations, train a policy, then watch that policy take the controls — with no GPU anywhere in it.
CPU training: 14 s on one core
in progress
case study →
SpeechLens
ML · CVA fully local language-ID and transcription pipeline — faster-whisper and Silero VAD behind an explicit anti-hallucination harness — running large-v3 at 3.9× realtime on a CPU with no GPU at all.
CPU throughput: 3.9× realtime (9950X, large-v3 int8)
in progress
case study →
CUDA MPPI Controller
Systems · HPCA sampling-based model-predictive controller whose CPU reference and CUDA kernel call the same rollout function, so they cannot disagree about the algorithm — verified on the CPU side, with the Jetson GPU run still to come.
cartpole swing-up, CPU: 99–100 of 100 seeds
in progress
case study →
Remote ML Compute over Tailscale
Systems · HPCA headless Ryzen 9 9950X reachable by name from anywhere over a private mesh — the CPU-only box that ran every experiment behind SpeechLens and SpeechForge.
unattended compute: ~15 h of studies in one session
in progress
case study →
SpeechForge
ML · CVZero-shot voice cloning inside an evaluation harness that scores every result against a ground-truth row — and caught seven of my own measurement errors, two of them already written up as findings.
WER from one input fix: 0.112 → 0.022 (n=60)
in progress
case study →