Recent Projects

AI-Driven Iterative Data Analysis with GPT Models

Employing OpenAI's GPT-3.5 and GPT-4 models, I built an autonomous data scientist capable of iterative hypothesis testing and data analysis using Python. The AI agent, designed with a scientific method approach, generates hypotheses based on user input and past experiment outcomes. After user's hypothesis selection, the AI constructs Python code to test the hypothesis and produces a report summarizing the findings. This report then informs the next set of hypotheses, effectively creating a continuous cycle of learning and analysis.

I leveraged best practices from OpenAI to improve bot prompting, which was particularly instrumental in overcoming initial challenges in generating executable code. This was achieved by incorporating specific developer environment information into the prompts. The project took inspiration from Smoldev and Langchain, utilizing their frameworks for managing autonomous language model agents.

The effectiveness of this AI agent is underscored by its contributions to several other projects detailed on this same page. This endeavor reflects my proficiency in applying machine learning and data science principles, tailoring autonomous agents, and overcoming unique challenges in innovative AI applications. 

Leveraging AI for Real-time Game Design Feedback

As a data scientist, I built a sentiment analysis pipeline specifically for live service game designers, employing Python and the Discord API. This pipeline batch processes Discord conversations daily, identifying key topics and emotions in player feedback.

The pipeline integrates ChatGPT 3.5 and 4 via API to decipher conversation themes and associated emotions. For instance, it can differentiate between a player expressing frustration at a game balance issue, and another player expressing excitement over the same issue, providing developers with nuanced insights into player sentiments. 

This pipeline offers a practical, repeatable method for gathering actionable player feedback that might otherwise be too time-consuming to collect. By graphing sentiment against time, it provides a clear depiction of player reactions to changes in gameplay. 

One challenge was devising a suitable output template for ChatGPT to ensure the insights delivered to game designers were clear and actionable. Developed during a contract role, this pipeline aids in automating the nuanced work of tracking and interpreting player sentiment, demonstrating my ability to meet complex project requirements.

Spatial Strategy in PUBG: Player Success Analysis 

In this data analytics project, I collected telemetry data from 200 PUBG matches using their Developer API. I analyzed player landing locations and distances from other players using Python and SQLite3. I handled data inconsistencies by normalizing player ranks in incomplete matches and excluding matches with bots.

My analytics-based approach revealed the importance of players' distance from opponents and identified south-east and north-west as advantageous landing locations. Insights were visualized using Matplotlib, highlighting patterns that can influence player strategy and game development.