Learning projects I run outside work, mostly where AI systems meet real hardware constraints. Each one ships code and its own measurements.
Edge Inference Benchmark Lab
A reusable benchmark harness for quantized inference, built around MobileNetV2 on an NVIDIA Jetson Orin Nano Developer Kit. It compares FP32, TF32, FP16 and INT8 in terms of latency, throughput, Top-1/Top-5 accuracy and energy per frame.
ONNX · TensorRT · CUDA Graph · NVIDIA Model Optimizer · tegrastats telemetry · Python
Read the write-up → · Repository →
Interruptible Voice Runtime
A push-to-talk voice command runtime on the same Jetson Orin Nano. Silero VAD detects the speech, faster-whisper transcribes it on the GPU, and a closed grammar turns the text into a JSON command with a latency breakdown. A single-writer, event-driven design with a generation counter guarantees that a cancelled utterance can never emit a stale command.
Silero VAD · faster-whisper · CTranslate2 · CUDA · PortAudio · Python