Nikhil A V

DriveLink

Vehicle-to-vehicle intelligence: cars that negotiate lane changes and merges with each other instead of guessing. Now being incorporated as DriveLink V2V Intelligence LLP.

• drivelink.tech runs a live V2V traffic simulation in the browser: car-following, negotiated lane changes, on-ramp merges.
• RandomForest models decide when a car should merge, change lanes or wait. They drive a Godot highway sim live over WebSockets, and the same models are validated inside the SUMO traffic simulator.
• DriveLink Connect: a Flutter app that lets anyone reach a car's owner by scanning a QR tag, without sharing phone numbers. Includes crash SOS.

Arvyo

When a star dims, something crossed it: a planet, an eclipsing binary, a blended background star, or a starspot. Arvyo rebuilds all four from physics, fits each to the TESS or Kepler light curve, and reports which one the evidence prefers, with the reason attached.

assay

An agent that investigates fraud alerts on a TigerGraph transaction graph and writes a case file, with the graph query behind every claim. Its model is trained only on the bank's closed cases and tested on a month it never saw: AUC 0.964 against the bank's own risk score at 0.866.

CHOWKI

A delay-tolerant network for mountain trail booths with no internet and no signal. Hikers carry ledger snapshots between booths over Bluetooth, acting as data mules. A local Gemma model parses incident reports and flags overdue hikers.

Drone detection

A YOLO11 drone detector I trained with an extra small-object detection head for distant targets. 0.915 mAP50 on validation and 131 FPS inference. It is the camera channel of AURA, a high-altitude counter-drone prototype.

flowtype

Press a hotkey, talk, and clean text appears at your cursor in any app. Speech recognition runs on the GPU with Whisper, a small local model removes filler words, and nothing leaves the laptop.

Friction Heatmap

A VS Code extension that measures which lines of code are hard to understand by watching how people hesitate: dwell time, revisits, pauses before editing. The question it tests: does human hesitation predict bugs better than code complexity metrics?