< 086fbc8c75 />

< 02db495216 />

< bf62e56bac />

< c8db44eac5 />

< 1b274ad93d />

< b0c91ac66a />

< b4a73a3e92 />

< f52ebfe281 />

< e171d941f0 />

< fb6d9bde65 />

< 1a385c0048 />

< b58e04a9b7 />

< 83fa785898 />

< 543696c485 />

< f25c4a4b21 />

< 0afd19592f />

< 347e4599be />

< 279b143230 />

< d3f5138020 />

< 198cafdbc5 />

< 775ccd7a4a />

< e59c3f1698 />

< 28b50f74eb />

< 9b59d6ace1 />

< c6644ceb11 />

< a8660beab7 />

< 65c0081784 />

< a6c7eed6c3 />

< 131c93a036 />

< 216acd2963 />

< 46848ee9e5 />

< ab14cb99f5 />

< b2fb3cbed6 />

< e9c197e05c />

< f6cb295be3 />

< be94d7433c />

< 98a21e0f4f />

< 4d07a3393b />

< cf098c3558 />

< 74831bed1a />

ALL PROJECTS

FEATURED

Vibe Coding Security Benchmark

Honors Thesis · AI-Generated Code Security

securityresearchai

Benchmarked 12 LLM-generated web applications across 5 attack scenarios using a custom Playwright-based exploit verification pipeline and CAF metric. Key finding: security failures concentrate at integration boundaries, not isolated logic bugs.

PythonPlaywrightFlaskExpressReact

> 12 apps · 5 attack types · published

FEATURED

YC API Penetration Test

Production Security Audit

security

Cold-emailed a YC founder, got pre-authorized API access, and confirmed a cross-tenant cache isolation vulnerability (CWE-200/208) reproducible 10/10 times. Full writeup published.

PythonBurp SuiteREST APIsCWE-200

> 10/10 confirmed · $300 bounty

FEATURED

SpectralQuant Security Benchmark

Original Research · KV Cache Attack Surface

securityresearchai

Identified that the static calibrated eigenbasis in SpectralQuant creates a predictable attack surface. Targeted eigenbasis perturbations outperformed random noise in 91.7% of cases across 28 layers of Qwen2.5-1.5B.

PythonPyTorchQwen2.5NumPy

> 91.7% attack success rate

ZkLoRA

Applied Cryptography · Verifiable AI Fine-Tuning

securityresearchai

Zero-knowledge proof system for verifiable LoRA fine-tuning -- cryptographic proof that AI model adaptation occurred correctly without revealing weights. Built 8 ZK backward pass gadgets at distilGPT2 scale.

C++emp-zkmlPythonPyTorch

> 8 ZK gadgets · distilGPT2 scale

gittuf

Open Source · Supply Chain Security

securityresearch

Active contributor to gittuf, a Git supply chain security framework built on TUF, incubating under the Linux Foundation OpenSSF. Fixed RSL synchronization bugs, annotation entry logic, and silent error swallowing in commit operations.

GoTUFGitOpenSSF

> Linux Foundation · OpenSSF

PratGPT

Multi-Agent Telegram Orchestrator

Multi-agent Telegram orchestrator with specialized agents for PhD application tracking (Gmail polling, Claude-drafted replies), content drafting, and budget management via Plaid. Event bus architecture with SQLite state.

PythonTelegramClaude APIPlaidSQLite

> Always-on · Railway deployed

CarbonCompass

HackHarvard 2024 Winner · Climate Analytics

Climate-impact analytics application using Cloudflare Workers and D1, integrating LLM-powered data extraction through Gemini and Amazon Q to deliver real-time carbon-footprint estimates for consumer products.

Cloudflare WorkersD1GeminiAmazon QPython

> HackHarvard 2024 Winner · <150ms

SecureChat

Encrypted Messaging & Exploit Simulation

security

Built and reverse-engineered a full-stack messaging platform using Flask, SQLite, and AES-ECB, simulating encrypted DHE-AES conversations. Launched XSS, ECB replay, and Diffie-Hellman key manipulation attacks using Selenium automation.

FlaskSQLiteAES-ECBSeleniumPython

> Full security simulation

Sherpa

Autonomous Vehicle Collaboration Platform

Data-sharing platform for autonomous vehicles enhancing traffic flow and reducing congestion using real-time communication and traffic APIs.

Next.jsTomTom APIYOLOVector SearchFlask

> 83.4% accuracy · 15% congestion reduction