Electronics & Communication Engineering undergraduate. I build systems where hardware, signals and machine learning meet — and I care about measuring things properly before claiming they work.
Heading toward VLSI and digital design, with a strong interest in running AI efficiently on constrained hardware.
JARVIS — a local-first voice assistant that runs entirely on a laptop with a 6 GB GPU. No cloud LLM.
- Voice: Whisper large-v3-turbo + Qwen3-4B on-device; time to first spoken word cut from 22–97 s to 0.6 s by diagnosing hidden reasoning tokens
- Memory: 110k messages indexed for episodic recall with hybrid BM25 + vector search and Reciprocal Rank Fusion; finds a single line from years back from a casual Hinglish question
- People model: tie-strength scoring validated against labelled data (95% precision@20), including a case where the data contradicted my own hypothesis
- Sensing: ESP32-S3 firmware for Wi-Fi RSSI/CSI presence detection, and an LD2450 mmWave radar decoder
- Honest by design: abstains instead of inventing memories it doesn't have
Every decision is documented with the measurement behind it in the engineering log.
Languages: Python · C++ (Arduino / ESP32) ML & retrieval: scikit-learn · PyTorch · sentence-transformers · LanceDB · BM25 LLMs & speech: Ollama · Whisper (faster-whisper) · local quantised models Hardware: ESP32-S3 · HLK-LD2450 mmWave radar · serial / WebSocket telemetry Engineering: Git · benchmarking · reproducible environments
- Taking JARVIS's presence sensing from simulation to real hardware
- Building toward digital design and FPGA work