An LLM-based agent that interprets natural-language privacy rules and autonomously enforces them across text, audio, video, and tabular data. It selects or synthesizes tools, executes them in a sandbox, verifies outputs, and replans when verification fails.

The system introduces dynamic tool generation, unified multi-modal policy handling, and a closed-loop verification + recovery pipeline with full auditability (manifests + logs).

Implemented GPT-2 (124M parameters) in PyTorch from scratch, replicating its tokenization, positional embeddings, and self-attention mechanisms. Achieved 26% HellaSwag accuracy (vs. OpenAI's 28.92%) with just 2 days of training on a single NVIDIA A600 GPU, compared to OpenAI's month-long cluster (NVIDIA V100s). This work demonstrates how strategic architectural choices can yield near-benchmark performance under resource constraints. Future work includes fine-tuning for code completion and IDE integration, leveraging these efficiency insights.

Engineered a full-stack data analysis platform capable of ingesting and modeling over 220,000 sports events using probabilistic methods. The core focus was developing a scalable, event-driven pipeline (Python/SQL) to find positive expected values opportunities in the sport markets.

Key AI and mathematical components include the application of the Shin method for precise inference of implied probabilities and the deployment of Monte Carlo simulations to generate thousands of scenario-based projections. This simulation environment rigorously models volatility (Standard Deviation) and capital growth (Bankroll Projections) to inform data-driven decision-making.

Designed and built a self-balancing robot using Lego Spike, starting with deriving equations of motion and formulating a state-space model. Validated the design through MATLAB simulations, static/dynamic sensor testing, and iterative hardware prototyping. Achieved close alignment between simulated predictions and physical performance, demonstrating robust control system integration.

Developed and trained a Deep Q-Learning agent to autonomously play Snake, demonstrating core RL principles: state-action mappings, reward function design, and ε-Greedy exploration. Achieved consistent scores of 60+ within 80 training episodes, showcasing rapid convergence in constrained environments.

Collaborated with a 4-member software team to integrate inertial sensors (MPU6050), gps sensors (Neo-6M), pressure sensors (MPL3115A2), and triple-axis magnometometer (MMC5603) into a student-built rocket. Developed firmware for Arduino Nano and STM32 to process real-time flight data, ensuring stable trajectory tracking. Contributed to the team’s goal of achieving precise altitude control.