Ritik Arorahi.

Most web vulnerability scanners are either dumb — spraying payloads and pattern-matching responses — or they need an expert driving. Crakrr is an attempt at a third option: a scanner that reasons about a target the way a human tester does.

It orchestrates Gemini and Claude models through a universal MCP server, letting AI agents crawl an application, form hypotheses about where XSS and SQLi might live, and then verify actual exploitability with Selenium-driven payloads instead of guessing from response strings. The MCP layer means any model that speaks the protocol can be swapped in as the brain.

The paper behind it won Best Paper at ICCDM-2026. The desktop client also has a custom OpenGL interface with hand-written shaders and physics-based animation, which the scanner absolutely did not need. It has them anyway.

A Formula Student EV generates a firehose of data and, historically, no good way to see any of it while the car is moving. We co-invented and patented a modular data-acquisition system to fix that.

The hardware is a custom PCB built around a Teensy 4.1, sampling the IMU, CAN bus, and analog sensors at 100Hz and streaming over MQTT. On the receiving end, a PyQt ground station renders live motor phasors (Id/Iq) and computes torque ripple in real time — so the drivetrain team can watch the motor mis-behave as it misbehaves, instead of reconstructing it from logs in the paddock afterwards.

It's the project where embedded C, network protocols, and desktop GUI work all had to hold hands. The patent office agreed it was novel, which was nice of them.

A quantum-safe, anonymous, terminal-based messenger — for people whose threat model includes adversaries that don't exist yet. "Just use Signal" is good advice. This is for when it isn't paranoid enough.

Key exchange is hybrid X25519 + ML-KEM-768, so breaking it requires beating both classical and post-quantum assumptions. Messages are sealed with ChaCha20-Poly1305 and routed over Tor. On exit, it wipes its own memory following DoD 5220.22-M — the messenger equivalent of eating the note.

Written in C++17, open-sourced under MIT. Building it meant learning exactly how many places sensitive bytes hide in a running process. The answer is: more.

Deepfake generators need clean source images. NoiseNet makes yours quietly poisonous: a defensive perturbation system that embeds adversarial noise into photos so that generation models choke on them, while humans see nothing wrong.

The perturbations live in DCT mid-frequencies — high enough to survive compression, low enough to stay invisible — and are hardened with Expectation-over-Transformation so they survive the crops, resizes, and re-encodes that images suffer in the wild.

Training runs on a parallel GPU + Intel NPU pipeline with asynchronous evaluation, which meaningfully accelerated convergence. Fighting AI with AI, with a straight face.