Rain Piao · Los Angeles
I’m a UCLA sophomore studying Statistics & Data Science and Applied Mathematics.
Right now, I’m exploring character diffusion with 字生, studying demonstration rankings with BruinML, and building a poker-action scorer with ACM AI.
Outside class, I make music, act with CFan, and beatbox with SouLA.
01 / Education
Where I’m
learning.Education
University of California,
Los Angeles
B.S.Statistics and Data Science
B.S.Applied Mathematics
Specialization in Computing · GPA 3.983 / 4.00
College Honors Program · Dean’s Honors List, Winter & Spring 2026
Selected coursework
Data Mining, Probability, Regression Analysis, Experimental Design, R Programming, C++ Programming, Linear Algebra, Discrete Structures, Multivariable Calculus, and Differential Equations.
Before UCLAThe Pennington School2022–2025
Work
Professional experience.
02 / Work
Selected work.Projects & research
ML projects
Watching noise turn into a character.Class-Conditional Character Diffusion — 字生
Working pilot · interactive demo
I trained a small conditional U-Net on 3,072 rendered glyphs. Pick a character and watch the reverse process in your browser.
- Implemented a class-conditional DDPM in PyTorch, using learned character embeddings and sinusoidal timestep embeddings; trained on 3,072 synthetic 48 × 48 glyphs across 12 labels and two font families.
- Used AdamW and EMA weights; built a seeded, checkpoint-backed sampler exposing 101 states of the original 1,000-step reverse chain. The website runs the same weights in-browser, with the original DDPM update and an optional faster DDIM sampler.
Scoring the available poker actions.Poker Decision Modeling
A MiniLM-based action scorer, plus model-engine integration for PocketSolver. My work covers candidate scoring, data audits, and scenario-grouped evaluation.
- Implemented a PyTorch decision model with a 33M-parameter MiniLM encoder, candidate-token pooling, masking, and a shared head to score variable action menus in one batched forward pass.
- Audited 563,200 PokerBench source records and evaluated a 10,000-example development pool with scenario-grouped five-fold cross-validation. MiniLM achieved 65.2% exact-action accuracy and 28.8% lower log loss than a linear baseline; the baseline had higher action accuracy.
- Contributed to the team’s model-engine wrapper, comparing local-model and hosted-LLM probabilities and response times through a common interface.
Research
How many rollouts make a ranking reliable?Sample Complexity of Demonstration Influence Ranking
Ongoing · learning theory & robotics
Studying how many evaluation rollouts are needed to identify the most useful robot demonstrations, and how estimator variance affects confidence in that ranking.
Finding unfamiliar regions in images.CLIP-Based Out-of-Distribution Detection
Completed · ALT 2027 submission under review
Three Monte Carlo studies of what makes small unfamiliar image regions hard to detect. I implemented a clipped scan and contributed experiments and figures to our ALT 2027 submission.
- Designed three Python Monte Carlo studies, including a 225-configuration sweep of noise distributions, object geometry, and signal strength.
- Implemented a clipped spatial scan that raised simulated detection power from 19.9% to 97.7% in one Gaussian-mixture setting at a 5% false-positive target.
- Used separate fitting, calibration, and evaluation samples; examined uncertainty with Wilson intervals and paired bootstrap comparisons, and produced three manuscript figures. The manuscript was submitted to ALT 2027.
Software projects
Turning sheet music into MIDI.PDF-to-MIDI Conversion — Pusheng
In development · recognition and evaluation
An ongoing PDF-to-MIDI project. I’m evaluating recognition engines and note-level errors before building the musician review interface.
- Developing a PDF-to-MIDI and MusicXML workflow, beginning with recognition-engine evaluation and note-level error analysis.
- The musician-facing score review interface is still in development; the repository is not public yet.
From point clouds to passing cars.LiDAR Traffic-Monitoring System
A C++ pipeline that groups LiDAR points into vehicles, tracks them across frames, and estimates traffic speed.
Competitions & distinctions
USACO Gold Division
Reached the Gold Division of the USA Computing Olympiad in 2023.
Building a trading strategy.Algorithmic Trading Strategies
Five rounds of trading bots: rolling regression, inventory-aware quotes, options, and basket spreads. I iterated 20+ backtest variants and learned to check signals against actual fills.
- Built event-driven Python strategies with per-product modules, persistent JSON state, rolling OLS/EMA and z-score signals, and inventory-skewed quotes.
- Extended the bot with Black–Scholes valuation, cross-strike volatility fitting, decaying counterparty-flow signals, and rolling pair/basket spreads.
- Compared fill costs, inventory paths, and product-level PnL across 20+ backtest variants. Qualified in the top 16% on a four-person team and completed all five rounds.
A dataset, five people, 48 hours.Healthcare Data Analysis
I coordinated a five-person team and worked on the patient-level stress analysis: joining records, building care-journey features, and turning the findings into a five-minute presentation.
- Joined survey, demographic, and encounter tables in R; aggregated repeated questionnaires into patient-level stress scores and derived encounter counts and gaps between distinct visit dates.
- Applied Spearman correlation with Benjamini–Hochberg correction and cohort tests; explored age-bin-adjusted logistic models of screening and social needs by patient-portal status.
- Coordinated five people through 48 hours of analysis. Built ggplot2/R Markdown reports and scripted a five-minute patient-journey presentation connecting four analyses.
03 / Toolkit
What I work with.Skills
Programming
Python
C++
R
Java
Implementation, data pipelines, and the small programs behind these projects.
- Algorithm design & problem solving
- Data pipelines & event-driven programs
- Input validation, batching & masking
Statistics & data
NumPy
pandas
SciPy
Matplotlib
ggplot2
From patient-level data joins to controlled Monte Carlo studies.
- dplyr, tidyr, data.table & R Markdown
- Spearman tests & Benjamini–Hochberg correction
- Regression, Wilson intervals & paired bootstrap
Machine learning
PyTorch
Hugging Face Transformers
scikit-learn
Conditional diffusion, action scoring, and experiments with frozen image features.
- DDPM / DDIM & conditional U-Nets
- Transformer encoders & candidate scoring
- Grouped cross-validation & baseline comparisons
Development & publishing
Git
Linux
Jupyter
LaTeX
Reusable checkpoints, recorded experiments, and the website you’re reading.
- Version control & experiment artifacts
- HTML, CSS, JavaScript & GitHub Pages
- Reproducible configurations & result records
Mathematical foundations
04 / Music
There’s music here, too.
Two of my earlier tracks.
01 / Original track
Breath Away
Melodic Dubstep
Yufei Piao
- Tempo
- 73 BPM
- Key
- B♭ major / G minor
02 / Original track
Camelia Overdose
UK Hardcore
Yufei Piao
- Tempo
- 235 BPM
- Key
- C minor / E♭ major
05 / Outside class
The rest of the week.
CFan Theatre
Stage credits · Fall 2025–Fall 2026
- 《井底天空》毕宁
- 《安可》Winter skitG
- 《墙》The WallNo. 7
- 《生死客》Fall playKline
Alongside acting, I worked on sound and wrote nine original cues for 《墙》. I also compose for 《生死客》.
CFan TheatreSouLA A Cappella
I’m on vocal percussion and tenor. I’ve also arranged for our media projects and worked on recording, mixing, and video.
SouLARunning
Jersey City Half Marathon
- Distance
- 21.25 km
- Workout time
- 1:46:48
- Avg. pace
- 5′02″ / km
View run record ↗
VALORANT
- Current rank
Platinum 3- Peak rank
Diamond 2
Last updated October 2026
Chess
Blitz on Lichess.
Lichess








