CV

You can download the full PDF version here: Jiayou Liang’s CV (PDF)

Education

• M.S. in Applied Data Science, University of Chicago, GPA: 4.0/4.0 (Oct 2024 – Dec 2025)
• B.S. in Financial Mathematics, First Class Honours, Hong Kong Baptist University, GPA: 3.77/4.0 (Sep 2020 – Jun 2024)
• Summer Course: Artificial Intelligence and Machine Learning (Grade: A), University of Oxford (Aug 2022)

Skills

• LLM & AI Algorithms: Fine-tuning, Transformer architectures, optimization algorithms, prompt engineering, inference optimization, vector databases (FAISS)
• Machine Learning: Deep learning, reinforcement learning, supervised / contrastive / self-supervised learning, computer vision, combinatorial optimization, integer programming
• AI Agent Development: LangChain, LangGraph, AutoGen, RAG (Retrieval-Augmented Generation), Azure Cognitive Search, multi-agent system design
• Programming: Python (PyTorch, scikit-learn, NumPy, Pandas), C++, SQL (PostgreSQL, MySQL), R, Bash
• Tools & Platforms: Git, Docker, Google Cloud, Azure, Tableau, LaTeX, Claude Code

Research Experience

• Apr 2025 – Present: Accelerating Large-Scale Adaptive Optimization via Reparameterization
  • University of Chicago | Advisor: Prof. Rebecca Willett
  • Algorithm design: Led the design of Auxiliary-Variable and Reduced EGOP algorithms to bypass the storage/compute bottlenecks of full-matrix EGOP reparameterization; theoretically tightened Adagrad/Adam convergence bounds by a factor of 1/d.
  • Low-rank + orthogonal complement: Used randomized SVD to extract the top-r eigenvectors of the EGOP matrix; introduced an orthogonal-complement auxiliary variable enabling lossless full-space optimization while storing only a small basis.
  • Empirical results: Implemented the module end-to-end and benchmarked on Fashion-MNIST, EMNIST, CIFAR-10; on a 78,400-dim layer, retaining <1.3% of components (r=1000) — and even r=50 — captured the loss landscape’s dominant geometry and accelerated convergence.
• Apr 2025 – Present: Neural Assortment Optimization (NAO)
  • University of Texas at Austin | Advisor: Prof. Rui Gao
  • Full-stack development: Built the NAO codebase from scratch, implementing an “extend–particle search–round” pipeline: Lovász tight extension lifts the discrete objective onto [0,1]ⁿ with exact optimal value, multi-particle projected noisy SGD with entropic-risk pooling explores the non-convex landscape, and chain rounding recovers a discrete assortment with no rounding-error term. Supports BAM, NL, and MMNL choice models.
  • Scaling & theory: Designed a rolling-window construction for capacity-constrained settings that maintains feasibility while preserving informative subgradients. Pushed tractable size to n=10,000 items at near-0% optimality gap on public benchmarks.
  • Beating baselines: Outperformed SOTA heuristics (e.g., ADXOpt) and neural baselines (NN/NNpp) on MMNL stress tests; at 10k+ scale, maintained fastest runtime and lowest error while exact solvers like Gurobi (conic integer formulation) hit OOM or exponential blow-up.
• Apr 2025 – Dec 2025: Preference Alignment of a Pathology Vision Foundation Model — Capstone | Honorable Mention
  • University of Chicago & Medical School | Advisors: Prof. Utku Pamuksuz, Dr. Samir Atiya
  • Vision-DPO architecture: First to transfer Direct Preference Optimization from LLM alignment to computational pathology. Designed a pseudo-probability framework (cosine similarity → temperature scaling γ=10 → sigmoid) converting TITAN encoder geometry into DPO preference likelihoods; LoRA (r=16) trains only 2.7% of base parameters with frozen reference model for implicit KL regularization.
  • Self-supervised preference data: Designed a label-free preference-pair pipeline: baseline classifier flags misclassified samples; UMAP + DBSCAN over correct clusters yields “preferred” embeddings vs. misclassified “rejected” targets. End-to-end: preprocessing → preference generation → LoRA fine-tuning → SCL → classification.
  • Results: Lifted PR-status AUROC from 79.75% → 81.92% and Cohen’s κ from 45.40% → 54.43% over raw-TITAN baselines; awarded Honorable Mention at the UChicago Applied Data Science Capstone Showcase.

Selected Publications

• Liu, R., Liang, J., Chen, H., Hu, Y. (2025). Analyst Reports and Stock Performance: Evidence from the Chinese Market. Asia-Pacific Financial Markets. — NLP, BERT, sentiment analysis
• DePavia, A., Cruzado, J., Liang, J., Charisopoulos, V., Willett, R. Faster Adaptive Optimization via Expected Gradient Outer Product Reparameterization. (Submitted to SIMODS) — Adam/Adagrad, low-rank preconditioning
• Atiya, S., Liang, J., et al. Validating Two Pathological Foundation Models for Breast Biomarker Detection. (In preparation)

Project Experience

• AI-Assisted Clinical Data Navigation Platform (LLM + Agentic AI) — St. Jude Children’s Research Hospital
  • Agentic NL-to-SQL interface: Designed and built an LLM agent chat interface enabling researchers to explore the Childhood Cancer Survivor Study (15,000+ de-identified variables across survivor and sibling cohorts) via natural-language questions, auto-translated to SQL with instant summary statistics and visualizations — dramatically lowering the barrier to pre-proposal data exploration.
  • RAG + multi-agent retrieval: Built a Retrieval-Augmented Generation pipeline integrating Azure Cognitive Search with a LangChain / LangGraph multi-agent framework for intelligent variable discovery, codebook lookup, and document retrieval, replacing manual hunting through legacy data dictionaries.
  • Production deployment: Delivered the end-to-end stack — secure PostgreSQL backend with role-based access control, LLM API integration, prompt orchestration, and automated data-processing flows — supporting both cohorts in a HIPAA-conscious clinical research environment.
• HikerScrolls — Outdoor Hiking Scrollytelling Plugin — Obsidian Community Plugin (Author)
  • Heterogeneous data fusion: Independently designed and shipped a published Obsidian community plugin parsing and fusing GPX track data, EXIF photo metadata, and 10+ map tile sources (Stadia, OpenStreetMap, etc.); rendered three visualization templates: scrollytelling narrative, photo scrapbook, and hand-drawn map.
  • Multimodal LLM integration: Integrated the Gemini LLM for AI trip-summary generation and AI-driven layout optimization; built a “Souvenir Store” module driven by multimodal prompt engineering to generate 5 categories of travel souvenirs (postcards, fridge magnets, etc.) from trip photos and metadata.
  • Full product surface: Architected the Journal Wizard, elevation-profile renderer, global Atlas map view, timeline sidebar, and bilingual layer; published to the official Obsidian community plugin store.
• Cross-Asset Order Flow Imbalance (OFI) Analysis — Quant Research Project
  • HFT feature engineering: Engineered 5-level Order Flow Imbalance features from high-frequency limit-order-book data on 5 cross-sector Nasdaq 100 stocks (AAPL, AMGN, TSLA, JPM, XOM); used PCA to aggregate the multi-level signals into a unified metric, raising the average self-impact regression R² from 0.54 to 0.65.
  • Cross-asset impact modeling: Built four contemporaneous regression families (PI/PII self-impact, CI/CII cross-impact) to quantify how a single stock’s OFI propagates to other names’ price movements; the cross-asset integrated CII model achieved best-in-class performance across all stocks (avg R² = 0.71, max 0.75 on AMGN), empirically validating order-flow information transmission in multi-asset pricing.

Internship Experience

• Jun 2023 – Aug 2023: Marketing & Sales Analyst Intern, Exion Asia (Huizhou) Co., Ltd.
  • Decomposed sales/revenue trends by region, industry, and product line; identified seasonal and promotional drivers for executive weekly and monthly reports.
  • Built multi-dimensional Tableau dashboards (KPI tracking, sales funnel, geographic distribution), automating the weekly/monthly reporting workflow.
• Jun 2022 – Jul 2022: Customer Service Manager Intern, Bank of Communications, Huizhou Branch
  • Rotated through teller and lobby positions serving 200+ customers with zero complaints; supported wealth managers in client needs analysis and product recommendation.

Awards & Honors

• Honorable Mention — University of Chicago Applied Data Science Capstone Showcase (2025)
• First-Class Scholarship — BNU-HKBU United International College (2021, 2022, 2023); President’s Honor Roll
• Third Prize — Guangdong Division, China Undergraduate Mathematical Contest in Modeling (2022)