• VelocitiPM is an AI-assisted product management platform: it helps a product manager take an idea from a rough strategy all the way to an engineering-ready ticket, without switching between five different tools to do it. I work full stack on it, across the web app, the AI agent backend, and the database layer.
• Built the step where a product manager turns a validated idea directly into a ready-to-hand-off engineering ticket, complete with a proper user story, in a single flow, replacing what used to be a manual rewrite into a separate ticketing tool.
• Built the piece that carries a feature's actual design and prototype context along with its generated requirements, so a PM (or the engineer picking it up next) can see what was really designed instead of hunting down a separate design link.
• Designed the automation that keeps a feature's status in sync with its parent roadmap item as work moves forward, so a PM's roadmap reflects real progress automatically instead of needing to be updated by hand.
• Rebuilt the first-time experience for a new product manager signing up, from registration through their first guided tour of the platform, to cut down on drop-off in the first few minutes.
• Also worked on the AI agent side that generates the initiatives, user stories, and requirement docs a PM works from, improving how reliably it produces usable output.
• Wrote automated tests alongside nearly every feature to keep the platform dependable as new capabilities shipped quickly.

• Designed and implemented a training-free activation steering method in PyTorch to improve CodeLLM reliability for automatic bug-fixing tasks.
• Computed a correctness vector from contrastive buggy/fixed code pairs and injected it into the model's residual stream during inference using PyTorch forward hooks.
• Boosted CodeLlama-7B repair accuracy from 48.12% to 56.25% (+8.13pp) and eliminated invalid code generation entirely (4.38% → 0.00%).
• Improved Qwen2.5-Coder-7B accuracy from 63.12% to 70.00% (+6.88pp) on HumanEval benchmarks.
• Achieved +4.43pp accuracy gain on Qwen2.5-Coder-14B across 565 real-world Java bugs in the Defects4J dataset.
• Ran 40+ experimental configurations across a layer-strength hyperparameter grid on HPC Slurm clusters.
• Novel finding: middle transformer layers are optimal for code generation tasks, while late layers are optimal for repair. That points to task-dependent reasoning stages in CodeLLM architecture.

• Engineered deep learning training loops inside Intel SGX hardware enclaves using C++ and the Darknet framework, isolating computation from OS-level vulnerabilities.
• Implemented AES-GCM encrypted model weight export/import, maintaining 100% data privacy even during simulated system compromise scenarios.
• Achieved ~50% training overhead (27.43s vs 41.07s) deemed acceptable for privacy-critical production workloads.
• Formally verified fairness properties of a 3-client federated learning protocol using NuSMV model checker and CTL temporal logic.
• Encoded and verified liveness, safety, starvation-freedom, deadlock-freedom, and bounded round completion properties across all aggregation rounds.

Focus on Artificial Intelligence, Machine Learning, and System Architecture. GPA: 3.5/4.0

Focus on Software Engineering, Data Structures, and Algorithms. GPA: 7.8/10