CV
My academic and professional CV.
Contact Information
| Name | Beixuan (Floria) Jin |
| Professional Title | Junior AI Researcher · Statistics @ UIUC |
| beixuan2@illinois.edu | |
| Location | Champaign, Illinois |
| Website | https://floriajin.github.io |
Professional Summary
B.S. in Statistics (Minor in Mathematics) at UIUC. I work at the intersection of statistics, machine learning, and AI safety—focusing on alignment, causal inference, Bayesian methods, and graph machine learning. My strength is cross-domain transfer: taking a framework apart functionally, testing causal directions with minimal experiments, and combining complementary streams into a working loop.
Experience
-
2025 - 2026 Research Assistant (Advisor: Prof. Xiaowei Luo)
City University of Hong Kong
Climate Risk Assessment System.
- Applied the ISI-MIP ATTRICI framework (Python, xarray, netCDF4, PyMC/SciPy) to generate counterfactual climate datasets, adjusting factual observations into stationary scenarios for aviation emission trade-off risk assessment.
- Ran the framework’s 8-step detrending and quantile-mapping pipeline to scale time-series data, fit variable-specific statistical models, estimate factual and counterfactual distributions, and preserve observational rank structure across transformed scenarios.
- Built probabilistic risk-analysis and visualization workflows to assess CO2 vs. non-CO2 aviation emission tradeoffs, using Python/R dashboards to summarize counterfactual climate shifts and scenario feasibility for research decision support.
-
2025 - 2026 Research Assistant (Advisor: Prof. Mi Xiang)
Shanghai Jiao Tong University
Aviation Personnel Mental Health Tracking Application.
- Implemented a session-memory component for an OpenCouch-style mental-health support app using LangGraph, FastAPI, and PostgreSQL/pgvector, with CoALA-inspired memory design supporting continuity across sessions and durable persistence.
- Structured memory into three CoALA-inspired layers—semantic facts, episodic arcs, and procedural rules—enabling the backend to retrieve user check-in summaries, support preferences, and recurring mental-health context before follow-up responses.
- Connected memory and safety logic across five supported interfaces—CLI, Next.js web chat, LiveKit web voice, Telegram DM gateway, and FastAPI backend API—to support longitudinal tracking of aviation personnel’s self-reported mental health status.
-
2025 - 2025 Quantitative Research Intern
Zhongshan Securities Co., Ltd
- Built a Python (pandas) ingestion and cleaning layer over 100+ listed companies’ H1 2025 filings, standardizing heterogeneous accounting line items into a unified schema supporting multi-dimensional comparison of revenue, profit, and gross margin.
- Integrated the Wind API for market data extraction (prices, PE ratios, ETF constituents, cash-flow series) and designed an indexed SQL schema over Wind-classified industry taxonomy, tuning composite indexes to the dominant query patterns of the research desk.
- Developed a Streamlit/Plotly dashboard tracking the top-10 companies by net asset value within each industry, with sector-level toggles supporting both single-sector drilldown and multi-sector horizontal comparison.
- Implemented walk-forward (rolling-origin) cross-validation for a volatility research effort—the correct validation scheme for time-series data where random k-fold splits leak future information and produce optimistically biased estimates.
- Authored an internal research primer on risk budgeting and risk parity, synthesizing Meketa and CAIA white papers, CFA Level III curriculum, and Bruder & Roncalli (2012); derived the equivalence between marginal and covariance-based views of risk contribution and quantified how volatility drag causes returns to leverage to saturate rather than scale linearly.
-
2025 - 2025 Fullstack Intern
PhinD LLC
- Adapted AWS’s open-source sagemaker-graph-fraud-detection solution, running its pandas-based preprocessing pipeline via SageMaker Processing to convert IEEE-CIS tabular transaction data into node features, labels, and relational edge lists for fraud classification.
- Evaluated graph-based fraud classifiers using DGL’s heterogeneous graph construction and RGCN training against two feature-based baselines (MLP and XGBoost) on SageMaker model training infrastructure.
- Deployed the end-to-end AWS workflow comprising two Lambda-triggered modules (data preprocessing and graph modeling), custom Docker processing images, S3 data channels, IAM roles, and CloudFormation templates for reproducible training and endpoint testing.
-
2024 - 2024 AI Software Developer Intern
NetEase Youdao Information Technology Co., Ltd.
- Built a multilingual OCR text recognition engine supporting 20 languages (Chinese, English, Japanese, Korean, etc.) at 96% per-image accuracy, 1.5s per image, and 100K+ daily volume.
- Developed OpenCV-based image enhancement for denoising, correction, and contrast adjustment, and applied YOLOv8 for text detection, achieving 94.2% localization accuracy and improving recognition success on blurred images by 30%.
- Built a multi-engine translation service with FastAPI, integrating Google Translate, Baidu, and DeepL APIs; supported 50+ languages at 300ms average response time and 500K+ daily translations with 92% accuracy.
- Developed an AI-powered text polishing system on OpenAI GPT-3.5 supporting academic and business styles, applying prompt engineering to improve readability by 30% and reach 89% user satisfaction.
Education
-
2023 - 2026 Champaign, Illinois, US
Publications
-
2026 Undergraduate Research Symposium 2026 (Submitted)
Beixuan Jin, Dylan Wu, et al. University of Illinois Urbana-Champaign.
Skills
Interests
Projects
-
MERINA — Improving Generalization for Neural Adaptive Video Streaming via Meta Reinforcement Learning
Research reproduction · Meta-Reinforcement Learning
- Reproduced a two-stage meta-RL pipeline for adaptive bitrate (ABR) streaming: VAE-based imitation learning for environment-context encoding followed by PPO-based policy optimization with GAE, on FCC and HSDPA bandwidth traces.
- Compared MERINA’s meta-adaptation generalization against rule-based baselines (BOLA, RobustMPC) on out-of-distribution traces (Puffer, Oboe), separating few-shot adaptation via latent-space task inference from zero-shot transfer robustness.
-
BLSP — Long-term, Medium Spatial Resolution Annual Land Surface Phenology with a Bayesian Hierarchical Model
Research reproduction · Bayesian Modeling
- Reproduced the BLSP hierarchical Bayesian model by configuring the rjags-to-JAGS Gibbs Sampler MCMC pipeline, fitting 6- and 7-parameter double-logistic curves to sparse Landsat EVI2 time series with credible intervals.
- Evaluated the trade-off between interannual variability fidelity and data-sparse robustness inherent to the hierarchical prior design, and built a Dockerized reproducible environment to resolve JAGS binary and R package version dependencies.
-
SecAlign — Defending Against Prompt Injection with Preference Optimization
Research reproduction · LLM Security & Alignment
- Ran SecAlign training with PyTorch, Transformers, DPO, and LoRA to teach the model to prefer legitimate instructions over injected prompts.
- Configured preference data construction with prompt-injected inputs, secure outputs, and insecure outputs to compare preference optimization against StruQ-style defensive SFT.
-
Circuit Breakers — Improving Alignment and Robustness with Circuit Breakers
Research reproduction · Representation Engineering
- Trained circuit-breaker adapters with CustomTrainer, compute_loss, and CircuitBreakerDataset using Transformers, PEFT LoRA, PyTorch, and DeepSpeed to alter harmful hidden representations while retaining benign behavior.
- Mapped the training workflow from retain data and circuit-breaker prompts through tokenization masks, hidden-state loss, and LoRA merge utilities to understand representation intervention versus refusal-only alignment.
-
Safe-RLHF — Safe Reinforcement Learning from Human Feedback
Research reproduction · LLM Safety Alignment
- Executed the full Safe-RLHF training sequence across SFT, reward model, cost model, and PPO-Lag scripts using DeepSpeed, Transformers, Hugging Face datasets, and W&B to study helpfulness and harmlessness as separate optimization signals.
- Inspected the framework modules under safe_rlhf.algorithms, scripts, and model-training entrypoints to connect actor model, reward model, cost model, KL control, and the Lagrangian safety constraint during policy optimization.
-
Latent Adversarial Training — Improving Robustness to Persistent Harmful Behaviors in LLMs
Research reproduction · AI Safety
- Traced latent adversarial attack and training workflows across latent_at modules and experiment notebooks using PyTorch, Transformers, PEFT, and TransformerLens to understand residual-stream perturbation for jailbreak, backdoor, and unlearning tasks.
- Adapted the experiment setup across notebooks, task installation scripts, and model-loading paths with Hugging Face, FastChat, DeepSpeed, and lm_eval to handle dependency compatibility and local runtime execution.
-
LogLLM — Log-based Anomaly Detection Using Large Language Models
Research reproduction · Anomaly Detection / System Security
- Reproduced the BERT, projector, and Llama classification pipeline on HDFS/BGL-style log sequences, aligning semantic log embeddings with decoder-based sequence classification for parser-free anomaly detection.
- Configured Python 3.8, CUDA 12.1, Torch 2.4, Hugging Face model paths, and dataset-specific sliding/session windows to resolve environment setup and large-log preprocessing issues.
-
Price Elasticity Estimation via Double Machine Learning
Personal project · Causal Inference — scikit-learn, pandas, RandomForest, OLS, parquet, Kaggle Online Retail
- Built a DML-based price elasticity pipeline with cross-fitting, engineered product/temporal/stock confounders, residualized log-price and log-quantity using RandomForest models, and estimated debiased causal elasticity via OLS.
-
Introspective Safety Alignment Pipeline
Personal project · Safety Alignment — vLLM, DPO, FastAPI, OpenAI API, DeepSpeed, Python
- Built a STAIR’s SI-MCTS safety-alignment workflow, tuning MCTS generation, actor prompts, model checkpoints, reward rules, and Python 3.11 dependencies to generate pairwise DPO data and compare CoT-style SFT with step-level DPO safety tuning.
-
Customer Incentive Recommendation System
Personal project · Causal Inference — DoWhy, EconML, causal-learn PC, NetworkX, MLflow, pygraphviz, Databricks CPU
- Customized a Databricks causal ML accelerator spanning causal discovery, effect identification, DML estimation, recommendation, and refutation, using background knowledge and graph corrections to estimate promotion effects from biased historical allocation data.