| Nebula AI |
Production-grade Agentic AI platform with DeepSeek, React, FastAPI, and PostgreSQL, containerized with Docker and fronted by NGINX for reliable deployment on Google Cloud Platform. Implemented custom MCP tools including bidirectional Google Calendar synchronization with conflict-free scheduling, and set up CI/CD pipelines to automate build, test, and deploy across web and mobile clients.
|
React, FastAPI, PostgreSQL, Docker, Nginx, CI/CD |
nebula-assist.com |
| GradieX |
Neural network framework from scratch in C++ — implementing forward and backpropagation, automatic gradient computation, and core tensor math with no external ML frameworks. Built modular layers and training infrastructure so arbitrary architectures can be stacked and swapped with minimal code changes.
|
C++, ML, Systems |
GitHub |
| prodbot |
Documentation agent for small teams. Runs as a git post-commit hook, reads the diff, writes changelogs, feature docs, and architecture diagrams. Routing is plain Python rather than model calls, so a typical commit costs only a few API calls. Output is staged and never auto-committed, which keeps a human review loop in place.
|
Python, Git Hooks, API Routing |
GitHub |
| connect4 |
Connect 4 agent trained with tabular Q-learning over 500,000 episodes and epsilon-greedy exploration. Q-table serialized to JSON so training resumes across sessions. Board logic and win detection were hand-rolled in C++, and the game plays against a human in the terminal with a full rules engine.
|
C++, RL, Q-Learning |
GitHub |
| ChopSticksAI |
Reinforcement learning agent for the Chopsticks game. Built a full rules engine, move generation, and Q-learning decision layer so the agent can play consistently strong games while remaining transparent and easy to debug.
|
Python, RL |
GitHub |
| GenreDetect |
End-to-end classification pipeline that infers music genre from structured audio features. Includes preprocessing, feature engineering, supervised training, and model evaluation with metrics-driven selection. The project is packaged so swapping the dataset or tuning model parameters is a configuration change rather than a rewrite.
|
Python, ML, Feature Engineering |
GitHub |