Electronics and communication engineering by training. My work sits at the intersection of machine learning and systems: model compression, inference runtimes, and computer vision applied to real measurement data.
I am principally interested in problems where a result must be both fast and correct, and where the two constraints are in tension.
neka-rt — A private, local-first inference engine.
semicon-driftsense — Localization of a high-resolution reference patch within a low-resolution SEM search frame under unknown zoom, unknown rotation, and possible absence. Submitted for the Applied Materials Drift-Sense challenge, SEMICON India 2026.
quantization-lab — Post-training quantization worked through from first principles.
| Machine learning | Post-training quantization, inference optimization, neural network internals |
| Computer vision | Template matching and localization under unknown pose; semiconductor metrology imagery |
| Systems | Local-first inference runtimes, backend services, command-line tooling |
Languages — Python, Java, TypeScript, C
Tools — PyTorch, NumPy, OpenCV, Docker, Linux, Git
| Project | Description |
|---|---|
| semicon-driftsense | Patch localization on repeating semiconductor layouts under unknown pose and possible absence |
| quantization-lab | Experiments in reducing model precision while preserving accuracy |
| SarvamBatchTranscriber | Command-line client for the Sarvam batch speech-to-text API |
| ForgeBoard | Collaborative project and task management backend, written in Java |
| karpathy-zero-to-hero | Neural networks constructed from a single neuron upward, implemented by hand |