Rohan Kaila
scroll to explore
“The only things I'm scared of is not being able to express myself and running out of Claude credits.”
Hi, I'm Rohan 👋
I'm a musician.
CS junior at Boston University. I build AI tools, automate workflows, and somehow ended up Security+ certified along the way.
Projects
Stopgate — Stop Watching Your Agent
This one's personal. I wanted to let Claude Code run unattended so I could actually walk away, but I didn't trust it not to quietly send my .env to an IP it read in a README, or delete the wrong thing. I'd just published AICES (proof that keyword harm-detection measures vocabulary, not intent), and Stopgate is me turning that into something I'd actually use. A local-first guardrail for AI coding agents (Claude Code, Cursor, MCP) that hard-stops the truly dangerous ones— secrets leaving the machine, destructive commands, spending — while auto-approving your normal work so you can walk away. The twist: on top of the usual command rules, it adds a provenance layer — it tracks where data came from and where it's going, so it catches a secret leaving even when the command looks innocent. On my own benchmark it cries wolf 2.75x less than a keyword gate at equal detection.
Refusal Circuits in Qwen2
A mechanistic interpretability study of Qwen2-1.5B: 7 experiments across 48 harmful/benign prompt pairs. I found the residual stream carries two geometrically distinct safety directions with opposite ablation effects — and that the intuitive one to monitor is the wrong one. Ablating the refusal direction across all 28 layers gets the model to generate non-refusal text.
AICES — Activation-Based Harm Monitors
It started as a question I couldn't get a clean answer to: instead of scanning text for bad words, can you detect malicious model behavior by reading the model's own internal activations — and does it actually work? A confound-controlled benchmark for prompt-level harm detection across five open-weight LLMs. The headline: every activation monitor scores 0.92-0.98 AUROC on the standard benchmark — but so does a bag-of-words classifier that never sees the model. I built a control set where that shortcut doesn't exist, and the monitor rankings changed.
Musicaguide — AI Assistant Inside the DAW
A JUCE/C++ VST3 plugin that loads into any DAW and answers mixing, mastering and sound-design questions about the track you're working on. The interesting part isn't the AI — it's the architecture: bring-your-own-key means it scales to unlimited users at zero hosting cost, and the user's API key never touches my servers or the repo.
Diet Decoder
A nutrition lookup tool that understands natural language. Describe food the way you'd say it, get back macros, micros, and calories. No exact USDA entry names required.
Zapier Automations
A growing collection of Zapier automations I've built for real people with real problems — AI-powered follow-up systems, smart outreach trackers, and more.
Legacy
AI-Powered Threat Detection in AWS
A production-grade AWS pipeline that trains Isolation Forest and Autoencoder models on VPC flow logs to detect anomalous network patterns from simulated red-team scenarios.
AI-Driven Chrome Extension
A Chrome extension that uses the OpenAI API and real-time DOM parsing to intelligently fill web forms, learning and improving from every manual correction.
Certifications
CompTIA Security+
CompTIA
Jul 2025
Designing & Building AI Products
MIT xPRO
Dec 2025
Google Cybersecurity
Google / Coursera
Sep 2024
About
Junior at Boston University studying CS, with minors in Business and Medical Anthropology. I've done cybersecurity work at Focus Technology, built out AWS infrastructure at Longterm Consultants (built that site too), and currently work as a campus ambassador for Perplexity AI.
Security+ certified. MIT xPRO in AI product design. I like building tools that fix the boring, repetitive parts of people's workflows. The stuff everyone complains about but nobody fixes. I also make music and fight fires on weekends, which my laptop has never appreciated.
Cool Firefighting Photos
Some moments from my time at the fire academy.
Contact
Open to research conversations and interesting projects.