Contributed to the multi-modal chat_template processing pipeline and video pre-processing pipeline. Enhanced test-suite for Vision Language Models (VLMs) including InternVL and QWEN2 family of VLMs. Improved the library's capabilities for handling video and image-text interactions.

SEED Emulator is a python framework that emulates operations of a real world internet. I am developing a module of the emulator which emulates oeprations of the SDN. In my implementation, I am using faucet (formarly known as RYU) as SDN controller and OpenVSwitch as virtual sdn enabled switch.

A framework to automatically test efficacy of security solutions proposed for IoT systems. With this framework researchers or developers of IoT security solutions can automatically generate test-cases and evaluate efficacy of their solutions and compare them with existing similar solutions intuitively. A regular user of IoT system can use VetIoT to pin-point security issues in the system and develop security policies for that.

Implemented embedding space attacks to bypass circuit breaker defense mechanisms in LLMs. Developed programs that find embedding matrices to trick fine-tuned LLMs into not recognizing harmful prompts, forcing them to generate harmful responses despite safety measures.

Broadcom Wifi/Bluetooth combo chip is used in many smartphones and laptops including iPhones, Samsung phones, and Macbooks. In this project, I emulated the firmware of the combo chip using qemu and fuzzed the firmware with randomly generated bluetooth packets. During fuzzing I found 2 CVEs: one crashing the firmware and another creating heap corruption error.

VPN client applications or built-in VPN configurators of popular OSs can leak data outside of VPN tunnel. In this project, I created a malicious wifi network. If a user connects to this malicious wifi access point, I can force user's traffic to follow a path outside of the VPN tunnel. In a more perilious scenario, this attack can be formed remotely by pairing with a dns hijacking

Built a tree based machine learning model to predict arrival status of flights arriving at SYR (Hancock Internation) airport from JFK (New York, NY), MCO (Orlando,FL), and ORD (Chicago, IL). To increase accuracy, we incorporated both historical weather data and flight arrival data from transtats. During prediction, we dynamically injected weather forcasting data to query.

Developed a Full-stack (front-end: React, Backend: Django) website with complete authentication and deployed it in AWS with apropriate scalability.