Khang Nguyen

M.S., Mathematical Data Science

Selected coursework: Introduction to Online Optimization (online learning and bandits), Mathematics of Machine Learning, Deep Learning and Its Applications, Mathematical Statistics for Data Science, Foundations of Mathematical Data Science

B.S., Data Theory. Dean's List (Fall 2021, Winter 2022, Winter 2023).

• First author on 4 papers and co-author on 6 in William Chang's group, with first-author work on mixed Nash equilibria of continuous games, corrupted MDPs, Lipschitz best-arm identification, and optimal gaze control.
• Led the UAI 2026 paper establishing tight rates for approximating mixed Nash equilibria of continuous games via entropy regularization.
• Act as the group's internal reviewer on nearly every paper: adversarially check proofs for gaps, rewrite flawed arguments and exposition, and run numerical experiments where the theory admits them.
• Lead author responses across the group's submissions, including two where reviewer scores rose after rebuttal.
• Lead the experiments team across concurrent projects, including a PPO baseline for autonomous driving on the Waymo Open Dataset via GPUDrive.

• Developed LSTM models for VN-Index stock prediction, testing 20+ technical indicators across 10 years of data.
• Applied two decomposition methods (EMD and wavelet transforms) to separate trend and cyclical components before forecasting.

Contract since Feb 2026; solo prototype Dec 2025 – Jan 2026.

• Sole AI engineer for a classroom-management LLM agent: 25+ function-calling tools cover the course lifecycle through chat, with citation-aware RAG and per-course student visibility filtering; now running on the production service ahead of market release.
• Designed the agent's action-gating model: every state-changing action requires an explicit tool call plus an out-of-band human confirmation the model cannot produce; nine write-action rules are codified in the spec, each enforced by a named regression test.
• Built defenses against adversarial student inputs (hidden-text prompt injection in submissions, jailbreak attempts), role-based tool dispatch with defense-in-depth re-checks, and fuzzy name resolution with confirmation prompts to prevent wrong-target actions.
• Maintain a 300+ test backend suite covering tool-call privacy gating, anonymity invariants, and security regressions.

Group project, lead engineer · yorru.net

• Built the full-stack platform with pgvector-backed RAG (OpenAI text-embedding-3-small) and a two-tier query system combining keyword matching with semantic search; the AI passively extracts guest preferences (dietary, budget) from group chat.
• Implemented real-time Socket.io chat across 3 interaction modes: 1-on-1 guest AI assistant, observer-mode group chat, and host dashboard.

Pitch prediction with LSTM and Conservative Q-Learning

• Built an offline RL pipeline for pitch selection from 8,400+ Statcast pitches (state: count plus last two pitch types; reward: negative change in run expectancy), sweeping CQL conservatism alpha and temporal reweighting with early stopping.
• Found that high-alpha configurations reported apparent policy improvements of up to 0.94 runs per pitch, several times the reward standard deviation of 0.20; cross-checked against training TD error and diagnosed the gain as Q-function miscalibration, since the metric is computed under the learned Q rather than a true value estimate.

Steam Games Insight Engine

• End-to-end NLP pipeline over scraped Steam reviews: KNN-based fake-review filtering, aspect-based sentiment analysis (NMF topic modeling + DistilBERT), and an ensemble of 3 fine-tuned transformers (GPT-2, RoBERTa, DeBERTa) with Jensen–Shannon disagreement analysis.

CUDA convolutional neural network

• CNN implemented from scratch in CUDA C/C++ (convolution, pooling, activation, and fully connected layers with forward and backward passes) with a CPU-vs-GPU benchmarking suite on MNIST.