During my internship hosted by Eric Malmi and Aliaksei Severyn, I played a significant role in a cutting-edge research project that enhanced language models (LMs) with search-based planning techniques to improve multi-step reasoning in board games such as Chess, Chess960, Connect Four, and Hex. My contributions were integral to the project’s development and success, culminating in a publication and integrating the model within Gemini as a Chess Gem.

My primary contribution focused on external Monte Carlo Tree Search (MCTS), where I addressed the challenge of balancing exploration and exploitation in low simulation count settings. I designed and implemented dynamic virtual counts, a novel mechanism that dynamically adjusts virtual count weights closer to leaf nodes, overcoming limitations of fixed virtual counts and virtual losses. This innovation significantly improved search efficiency and enabled the agent to achieve Grandmaster-level performance in Chess with a search move count per decision comparable to human players.

Additionally, I collaborated on:

• Pre-training a multi-action-value (MAV) Transformer model to capture value and transition functions across multiple games.
• Leveraging MAV within external MCTS as a policy and value predictor, improving state accuracy and minimizing hallucinations.
• Developing internal search, where the LM infers the search procedure by generating a linearized tree of potential futures, enabling performance scaling with an increased search budget.

This internship experience highlighted my ability to tackle complex technical challenges and contribute to impactful research, demonstrating the potential of combining LMs with search-based methods for broader inference and training applications.

Lead a team of five data scientists.

Behavioral Credit Scoring:

• Built a PySpark gradient-boosting model to predict consumer default risk probability, achieving a market-leading Gini metric results of up to 75%.

ML-Driven Marketing Campaign:

• Devised a data-driven campaign for promoting a heart disease drug to doctors on behalf of a top pharmaceutical company, which led to a 10% sales increase during A/B testing.
• Statistical analysis was conducted using Statsmodels, SciPy, and Python plotting packages.

Personalized Newsletter and E-Commerce Recommender Systems:

• Constructed a hybrid recommender system combining content-based and collaborative filtering, which achieved a 1.5% click-through rate during the proof-of-concept phase.
• Utilized Databricks, Python, PyTorch, and AWS in the technology stack.

Delivery Delay Estimation:

• Developed a customer support system for a shopping mall during the COVID-19 pandemic, which predicted delivery delays using a time-series ARIMA model supplemented with supervised learning techniques.
• The technology stack comprised Pandas, NumPy, and Sklearn.

Systemic Risk Model Execution Efficiency:

• Created a parallel version of a hill climber heuristic that made the optimization problem practically tractable. The heuristic’s runtime was limited to 3 minutes and, on average, generated solutions within 5% of the optimum, with the reported worst-case being 15% for tractable test-set instances.
• Employed a technology stack encompassed Python, CPLEX, and OR-Tools.

Treasury Department Cash Traceability:

• Constructed an uncollateralized debt tracking system by amalgamating diverse daily feeds to generate comprehensive firm-wide reports within seconds.
• Employed Q/kdb+, Python, PyQ kernel, and SQL for the development.

E-Trading Execution Limits Calibration:

• Fine-tuned an in-house model to prevent real-time executions during high-risk scenarios, employing a statistical analysis of e-trading clients.
• Utilized Pandas for the calibration process.

Listed derivatives liquidity:

• Created proof-of-concept for data-driven liquidation model based on intraday futures trading.
• Utilized Q/kdb+ for data processing.