Research Lab · Dan Nguyen

Saint Petersburg AI Agent

Aug 2025 – Dec 2025

Abstract

Can learned evaluators let Monte-Carlo Tree Search play a game with a large branching factor, and what do simulated games reveal about its balance?

Findings

MCTS with UCT (C = 2.0) and 4-ply early-terminated rollouts handled the large branching factor; the agent won 1st place in a class tournament.
Logistic-regression and random-forest win predictors over 56 engineered state features served as the learned evaluation.
A CatBoost model of rounds remaining cut learn-set RMSE from 2.20 to 0.62.
About 10,000 simulated games identified over- and under-powered cards, summarized in a balance-change report.

Methods

MCTS
UCT
Logistic regression
Random forest
CatBoost
Monte Carlo simulation

Full debrief →Paper ↗Code ↗

3-Row Chomp: Making Byrnes Effective

Aug 2025 – Present

Abstract

Byrnes (2003) proved that in 3-row Chomp the losing positions of each row r eventually repeat with a fixed period, but his proof gives no bound on when the repetition starts. Can that preperiod N(r) be bounded explicitly?

Findings

A streaming C++ solver computes exact losing positions for every row up to r = 50,000, validated against Brouwer's tables, Nivasch's census, OEIS, and a brute-force game tree.
Numerically, N(r) − √2·r stays within [−3.16, 5.70] for all r ≤ 50,000, sharpening the working conjecture to N(r) = √2·r + O(1).
The census turned up periods 6 and 8 beyond Nivasch's published set {2, 3, 4, 9}, first appearing at r = 11,036 and r = 18,718.
The two-pass referee gate has accepted an explicit computable bound, N(r) ≤ r + 2^O(r² log r), plus N(r) ≤ 3r for "stale" rows. Machine-refereed, not yet peer-reviewed.
From the 2025 capstone: proved by induction two infinite families of losing positions, (a+4, a+2, 2) and (a+11, a+7, 5), and conjectured the Gap Principle.

Implications

Next is a single human-checked write-up of the bound. Proving the "locking" lemma, which holds numerically up to r = 100,000, would give a linear bound N(r) ≤ 6r.

How it runs

PIPELINE.EXE

Solver
Streaming C++ census of losing positions, cross-checked against the published literature.
Explorer
MiMo-V2.6-Pro (xhigh reasoning) works one of two research islands, computing before it theorizes.
Ledger
Every claim is logged with its evidence; failed ideas go to a dead-ends log.
Referee
Two blind MiMo passes (temperature 0.3 and 0.8) see only the restated claim and a sandbox.
Split or reject → back to the explorer
Proven
Only an agreed, finished accept promotes a claim.
Operator
Claude Opus and I run novelty checks against prior art and read a daily digest.

Methods

Combinatorial Game Theory
LLM Agents
Adversarial LLM referee
C++
Python
GitHub Actions
OpenRouter
Proof by induction

Full debrief →Code ↗Paper ↗

AI Engineer Intern

KPMG · Berlin, DE · Jun 2025 – Aug 2025

Question

How can LLM-generated question-answer data for an audit-expert model be validated before it is used for fine-tuning?

Findings

Built a multi-threaded LLM-as-a-judge framework that validates the answer quality of every generated QA pair before fine-tuning.
Produced 500+ judge-validated QA pairs from a vector-graph RAG pipeline (LangGraph, Neo4j).
Cut inference cost ~17x by routing bulk generation to GPT-4o-mini with step-decomposed prompts and Pydantic-validated outputs.
Documented preprocessing as 5 notebooks sharing one structural chunking algorithm for both SFT and CPT pipelines.

Methods

Vector-graph RAG
LLM-as-a-judge
Chain-of-thought prompting
Structured outputs
Fine-tuning data preparation

Data Scientist Intern

KPMG · Berlin, DE · May 2024 – Aug 2024

Question

Which clustering approach best separates audit risk, and does it hold up under expert review?

Findings

Compared 5 clustering approaches across 15 sprints and selected the best; accuracy was validated with certified auditors.
A stochastic algorithm on top of a rule-based baseline improved cohesion and separation by 15%.
KNN imputation closed 30% ERP data gaps, enabling risk modeling on 500+ failed banks across two decades.

Methods

Unsupervised Learning
Clustering
KNN imputation
Stochastic algorithms
Expert validation

Healthcare Access Disparities Analysis

Jan 2024 – May 2024

Abstract

How unequal is healthcare access across Pennsylvania counties, and which gaps track worse outcomes?

Findings

Mental-health providers showed the highest inequity (Gini 0.185, Atkinson 0.4399).
Mental-health provider density correlated significantly and negatively with suicide rates in rural PA.
The regression explained 62.2% of the variance; rural status and lack of insurance were significant negative predictors.

Methods

Gini coefficient
Atkinson index
Welfare function
Regression
Econometrics

Full debrief →Paper ↗

Dan Nguyen

Research interests

Research

Saint Petersburg AI Agent

3-Row Chomp: Making Byrnes Effective

AI Engineer Intern

Data Scientist Intern

Healthcare Access Disparities Analysis

Methods

Poster wall

Saint Petersburg AI Agent

Abstract

Findings

Methods

3-Row Chomp: Making Byrnes Effective

Abstract

Findings

Implications

How it runs

Methods

AI Engineer Intern

Question

Findings

Methods

Data Scientist Intern

Question

Findings

Methods

Healthcare Access Disparities Analysis

Abstract

Findings

Methods

Whiteboard

Lab bench

🗺️ Town map

Around town

Welcome to my little world!

CRITTERDEX

TRAINER CARD