Projects

Adaptive Chess Puzzle Trainer

Type: Berkeley Chess School
Focus: Adaptive Difficulty
Technology: PyTorch, TensorFlow, Python, HTML/CSS, JavaScript, Git, Deep Learning, Neural Networks, Software Design, Cloudflare R2, Postgres, Alembic, React

ML NLP APP

Project Overview

Description: Intelligent trainer using Elo-based scaling for personalized puzzles.

Theory: Bayesian Elo, NLP for move annotations.

Results: Prototype with real-time feedback loop.

RL Speedrunner

Type: Personal Project
Focus: Training Optimal RL agent
Technology: Python, RL frameworks, Emulator Programming

RL ML APP

Project Overview

Description: Optimal agent utilizing RL to solve best path problems, with constraints.

Theory: Reinforcement Learning (Q-learning, exploration-exploitation analysis).

Results: TBD

Qualtrics Automation Workflow

Type: Haas School of Business
Focus: Reducing manual effort, automating mechanical processes
Technology: Python, JavaScript, HTML/CSS, YAML, Qualtrics, pandas, playwright, SMTP, tkinter, Jinja, pyinstaller, Git, Jaccard Similarity Algorithm, Windows, Mac, Software Design, Agile Development

STAT APP

Project Overview

Description: Built data pipeline framework: Qualtrics data sourcing, cleansing, synthesizing, pdf generation, email distribution.

Theory: Data piplining, Heuristic Graph-Based Clustering, Jaccard Support.

Results: Reduced several hours of work time, saved over $10K quarterly.

Rubriware - NLP Scoring System

Type: Haas School of Business
Focus: Reducing manual effort, automating intensive, mechanical processes
Technology: Python, PyTorch, PostgreSQL, Cloudflare R2, Pandas, FastAPI

DB APP NLP STAT

Project Overview

Description: Built an Agentic AI Agent to automate intesive manual classification work, outputs confidence levels for each score given for human oversight (High Confidence, Medium Confidence, Low Confidence).

Theory: Agentic AI, NLP Design, Database Design, Estimating Confidence intervals.

Results: From approximately 3 hours of manual work to 30 minutes of oversight, achieved ~90% accuracy on high confidence classification.