Machine Learning Engineer working at the intersection of large language models, multimodal reasoning, and quantitative finance.
I am a machine learning engineer and graduate researcher at NYU, where I study large language models, reinforcement learning from human feedback, and multimodal reasoning. My work has been published at ACL and EMNLP, and recognized with the AAAI Workshop Best Poster Award.
Before academia I spent three summers deploying production ML systems inside top-tier Chinese investment banks — building RAG pipelines over analyst reports, multi-modal alpha signals from financial text, and learning-to-rank systems serving live trading decisions.
I care about systems that actually work at the boundary of research and production: low-latency, auditable, aligned.
Architected a production RAG system serving GPT-4 QA/summarization over 10K+ analyst reports and Fed statements at <100ms p99 — reducing analyst manual review by 60%. Benchmarked 15+ forecasting architectures (TFT / LSTM / XGBoost / LightGBM), achieving 15% RMSE reduction on out-of-sample data. Scaled serving infra to 300+ QPS with MLflow-tracked A/B testing and automated retraining.
Independently developed a multi-modal market signal system — fine-tuned FinBERT-large on a proprietary Chinese financial corpus and fused embeddings with microstructure features via gated cross-attention, reaching 71% directional accuracy (+9pp over baseline). Built a NODE + TabTransformer stacked ensemble processing 1M+ daily predictions across 20+ contracts at <50ms p99, with β-VAE feature compression and purged k-fold CV.
Engineered a production-grade Airflow ETL for 4,000+ healthcare companies with Great Expectations + isolation-forest anomaly detection, cutting daily data cleaning from 8 hours to 45 minutes. Designed a 50-factor LambdaMART learning-to-rank stock screener with 85% top-decile precision over a 3-year walk-forward backtest, plus an IC-decay / mRMR / Shapley factor-quality framework for regime robustness.
Systematic analysis of how specialized multimodal architectures with spatio-temporal modeling enable reliable reasoning over musical performances — establishing design principles for audio–visual QA beyond generic multimodal baselines.
Proposed MSM-BD, a multimodal bot detection framework integrating profile images, tweet sequences, and user metadata through modality-specialized encoders and a novel Cross-Modal Residual Cross-Attention (CMRCA) fusion module — achieving state-of-the-art accuracy and F1 on TwiBot-22.
A specialized framework for audio–visual modeling in music understanding, addressing underexplored multimodal interactions, distinctive musical characteristics, and temporal alignment. Introduced annotated rhythmic and source features; achieved state-of-the-art on Music-AVQA 1.0 and 2.0.
Implemented a decoder-only transformer in PyTorch without reference to existing codebases — custom BPE tokenizer, multi-head causal attention with RoPE, RMSNorm, SwiGLU (LLaMA-3 parity). Scaled to 125M params via DDP + BF16, then built a full GRPO alignment pipeline with rule- and model-based reward functions, replicating the core mechanics of DeepSeek-R1 at small scale.
Test-time compute scaling: parallel generator ensemble (Llama-3 + GPT-4) producing N candidates under temperature-diverse sampling, supervised by a Process Reward Model scoring intermediate reasoning steps via Monte Carlo rollouts — replicating OpenAI’s PRM800K methodology. A meta-judge layer audits verifier reliability using structured LLM-as-judge rubrics.
Coursework — Machine Learning, Deep Learning, Realtime & Big Data Analytics, Big Data Application Development, Fundamental Algorithms, Database Systems.
Honour — one of only two international undergraduate students awarded the Highest-level Merit-based Scholarship. Coursework — Probability, Distribution Theory & Inference, Regression & GLM, Stochastic Processes, Computational Methods, Operations Research.