Hi, I'm
I build AI systems that ship — from retrieval-augmented agents and LLM fine-tuning to distributed training pipelines and autonomous robotics. Currently a Product Manager at Pegasystems and a Computer Science master's student at the University of Pennsylvania, with an Active Secret clearance. I love turning the world's hardest problems into working software, and I'm always open to connecting on ambitious AI and engineering work.
I'm a Computer Science master's student at the University of Pennsylvania who builds AI systems end-to-end. Right now I'm a Product Manager at Pegasystems, where I shipped a retrieval-augmented generation solution that cut incorrect AI responses by 20% and am designing a self-improving agentic loop that evaluates and refines its own outputs.
My path into AI runs through hands-on engineering. At General Dynamics Mission Systems I built automated pipelines that parse C++ and generate unit test plans, eliminating ~80% of manual test-planning effort. As an ML research assistant at Brandeis I reproduced methods from research papers, built distributed training across 8 TPU cores, and traced parameter evolution to find the tiny embedding layers that make or break a model. Earlier, at Arcada, I scaled a platform to 100K+ users and helped harden it against bot attacks.
I am a U.S. citizen with an Active Secret Security Clearance, based in Waltham, MA, and open to relocating anywhere in the United States. I love combining research depth with shipping instincts to solve hard problems. Outside of work I'm into nutrition and bodybuilding, basketball, reading, and chess.
University of Pennsylvania
MASc. Computer Science · Exp. 2027
University of Massachusetts
B.S. Mechanical Engineering · 2024
Waltham, MA
Active Secret
Autonomous RC car built on a Raspberry Pi 5 with a PCA9685 for precise PWM control of motors and servos. A PyTorch CNN maps raw camera frames to continuous steering and throttle outputs, deployed for real-time inference on edge hardware. Camera capture, inference, and actuation run in a closed-loop control system at 15+ Hz with sub-220ms end-to-end latency.
Comparative study of FFT, LoRA, and MFT methods on BERT-base model. Achieved 4.24% accuracy improvement on under-trained models using mask fine-tuning from 2025 ICLR paper while training only 0.037% of parameters.
I'm always happy to connect about AI/ML engineering, applied research, or building ambitious products. Whether you want to discuss collaborations, opportunities, or just trade ideas, feel free to reach out.