Custom Transformer Language Model

Python, PyTorch

Generated coherent, contextually consistent text outputs with improved semantic representation through the implementation of a 24-layer transformer language model and custom inference engine using 16 attention heads, 1024-dimensional embeddings, a 50,000-token custom tokenizer, temperature scaling, and top-k sampling.

Achieved stable model convergence for long-context learning over extended multi-day runs through the engineering of a fully optimized training pipeline on a 15GB text dataset using an RTX 4060 GPU, AdamW optimization, cross-entropy loss, and sequence truncation strategies.

Gaming AI Agent

Python, PyTorch, NumPy, Mathplotlib

Accelerated training throughput by ~2.5x and reduced end-to-end preprocessing latency from 120 ms to 40 ms per frame through the construction of a GPU-optimized tensorization pipeline using multiprocessing, parallel worker execution, and batching strategies.

Improved temporal stability and reduced prediction noise during gameplay imitation through the engineering of a real-time data acquisition system for supervised policy training using synchronized 30 FPS RGB frame capture from a live TF2 environment, temporal frame stacking (2.5x context window), and expanded action-space encoding.

Spotify API Multiplayer Web Game

HTML, CSS, JavaScript

Decreased client-side render jitter by 25% and maintained low-latency synchronization for up to 25 concurrent users by engineering a real-time multiplayer backend, integrating Spotify’s REST API with a custom WebSocket architecture.

Secured consistent cross-device rendering performance and stable real-time UI behavior by optimizing client update loops and DOM interactions with advanced event batching methodologies.

KSP KerboScript Boostback Landing

kOS, KerboScript

Improved trajectory efficiency and enabled precise return-to-launch-site maneuvers through the engineering of a comprehensive Kerboscript-based autopilot system using telemetry-driven ascent logic and real-time trajectory-prediction algorithms.

Achieved pinpoint landing accuracy to within ±0.25 m of target coordinates through the implementation of dynamic atmospheric re-entry and guidance systems using real-time pitch/yaw error corrections and an on-the-fly "suicide burn" throttle-control loop.

Autonomous Vehicle Autopilot

Lua

Engineered a modular autopilot control system implementing real-time navigation logic for ascent, cruise, and landing phases in a simulated environment.

Designed sensor-driven feedback loops for obstacle avoidance and trajectory correction, enabling adaptive real-time path planning.

Increased control-loop execution frequency from 15 Hz to 60 Hz by optimizing update logic and reducing per-cycle computational overhead.

Built reusable control modules for throttle, steering, and waypoint navigation, enabling scalable reuse across vehicle configurations.

Articulating Robotic Arm

Python

Engineered a 3-DOF robotic arm with MG996R servos and a PCA9685 PWM driver, using embedded systems programming on a Raspberry Pi Pico.

Increased servo response consistency by ~15% using optimized PWM timing.

Designed and 3D-modeled all mechanical components, applying precision tolerance control and iterative prototyping for smooth servo actuation.

Integrated manual and automated command modes, supporting scripted movements and real-time control input for flexible operational scenarios.

Secure Banking CLI

Python

Designed a secure CLI-based banking system with separation of authentication, transaction processing, and persistence layers.

Implemented salted SHA-256 password hashing with login throttling and verification safeguards to strengthen account security.

Designed a precision-safe financial transaction engine using Python Decimal, eliminating floating-point rounding errors in monetary operations.

Built JSON-based persistent storage with structured transaction logging and ISO-8601 timestamps for auditability and recovery.