Mechatronics Engineer — PSG Institute of Advanced Studies, Coimbatore
Building at the intersection of computer vision, edge AI, and industrial automation.

Pepper Studio (public)
Standalone, cross-platform sandbox for the Pepper 1.8 robot — drive it by hand, test your own code against an API-identical HTTP bridge, and drop in an optional AI brain (a cloud key, a local server you run, or a GGUF the app downloads and launches for you). Ships as a self-contained binary for Linux / macOS / Windows via GitHub Releases — no Python or Node install, no cloud required, works fully offline with no AI at all. Includes a 3D simulator, manual pose / speech / animation controls, a Sim / Real toggle, ⌘K command palette, first-run onboarding, and in-app voice + web search.

Project Prometheus — JARVIS v7 (private)
Privacy-first local AI assistant, clean-rewritten from 10,000 lines across 30+ files down to ~1,200 lines in 12 files. Qwen3.5-4B on llama.cpp with native function calling (24 tools), agentic tool loop (up to 8 rounds), LLM-based intent routing (replaced 60+ regex patterns), and a Karpathy-style auto-maintained knowledge wiki. SearXNG web search, Spotify control, Telegram bot, Kokoro TTS + Whisper STT voice I/O, SQLite memory, observation masking for context management. 32K context with ~31K usable for conversation. Runs on an RTX 3050 with a damaged VRM at 26W TDP — every token counts.

Weld Defect Detection — Industrial Vision System (Completed — May 2026)
Government-funded research internship under the Ministry of Heavy Industries, India. Sole ownership of the computer vision pipeline for retrofitting industrial welding arms with real-time defect detection. YOLOv8 on NVIDIA Jetson AGX Orin with TensorRT, GPIO relay logic for automatic arm and welder shutdown on detection. Integrated with STM32 controllers and CAN bus hardware. Built and shipped.

Gesture Media Control — Hand Tracking Input System
MediaPipe gesture recognition for media control and full mouse emulation on Linux. One-Euro adaptive filter for jitter-free cursor tracking, air-tap clicks, pinch-scroll with momentum, MQTT publishing for IoT integration. Single-file, ~730 lines, runs on any webcam. No cloud, no proprietary SDK.

Vision        →  YOLO26n · YOLOv8 · OpenCV · TensorRT · InsightFace · MediaPipe · Roboflow
AI / LLM      →  llama.cpp · Qwen3.5 · RAG · faster-whisper · Kokoro TTS · sentence-transformers
Robotics      →  NAOqi 2.5 · py_trees · Pepper 1.8 · Silero VAD
Languages     →  Python · C · C++ · MATLAB
Embedded      →  Jetson AGX Orin · STM32 · Arduino · CAN Bus · Keil uVision
Automation    →  Siemens TIA Portal · WinCC HMI · FESTO FluidSIM
CAD           →  SolidWorks · Autodesk Fusion 360
Systems       →  Arch Linux · Hyprland · Git · Docker · SSH

• B.E. Mechatronics Engineering — PSG Institute of Advanced Studies (2022–2026)
• International Semester — Leeds Beckett University, UK (October 2026)
• Research Intern — PSG-MHI Industry Accelerator Project, Ministry of Heavy Industries, Government of India (completed)
• IELTS Academic — Band 7
• ICAARS 2025 — Workshop on Generative and Edge AI for Automation
• Long-term direction: ADAS, vehicle development, and automotive embedded systems

Currently: building things that run locally.