Custom Transformer Language Model

Python, PyTorch

Developed a custom transformer-based language model from scratch, implementing a sophisticated architecture with 24 layers, 16 attention heads, and 1024-dimensional embeddings for robust text generation.

Engineered a custom tokenizer that efficiently processes and encodes text data, handling vocabulary sizes up to 50,000 tokens while maintaining semantic relationships in the input space.

Implemented an advanced training pipeline that processes raw text data, handles sequence truncation, and employs AdamW optimization with cross-entropy loss for effective model convergence.

Created an interactive inference system with temperature-controlled sampling and top-k filtering, enabling dynamic text generation with adjustable creativity levels and coherence.

Spotify API Based Multiplayer Web Game

HTML, CSS, JavaScript

Engineered a real-time multiplayer game leveraging Spotify's API, retrieving track and playlist metadata, optimizing for low latency and high concurrency.

Implemented a robust, WebSocket-based communication protocol to synchronize game state in real time across multiple concurrent users, ensuring seamless and reliable multiplayer interactions.

Developed a responsive, mobile-friendly front end using HTML, CSS, and JavaScript, employing adaptive design patterns to guarantee cross-platform compatibility and optimal performance.

KSP KerboScript Boostback Landing

KOS

Engineered a comprehensive Kerboscript-based autopilot system for Kerbal Space Program that automates flight phases from ascent to landing.

Designed an ascent algorithm that dynamically adjusts throttle and pitch based on real-time telemetry, calculating stopping distances and ensuring precise target apoapsis achievement.

Implemented a boost-back burn routine that leverages computed impact positions and trajectory data to synchronize thrust and steering commands for an accurate return-to-launch-site maneuver.

Developed an entry burn sequence paired with an aerodynamic guidance system, employing real-time pitch and yaw error corrections to align the vehicle with predetermined landing coordinates during atmospheric re-entry.

Engineered a suicide burn landing procedure that computes stopping distances on-the-fly, modulating throttle to ensure a controlled landing.

Autopilot Software

Lua

Engineered an advanced raycasting framework to continuously detect obstacles, track lane positions, and map intersections, ensuring the vehicle adapts to its surroundings in real time.

Designed a robust lane-following algorithm employing a PID controller that computes and adjusts steering corrections based on real-time telemetry, enabling smooth and responsive navigation.

Implemented intelligent decision-making routines for intersection navigation, complete with turn signal management and adaptive braking to ensure seamless transitions between maneuvers.

Optimized for low-latency multiplayer interactions, the system smoothly toggles between manual and automated control, providing both realistic driving simulation and intuitive user override capabilities.