Machine Learning · Computer Vision · Evaluation · Python

I ship models and measure them.

I'm Yougi, a data science student at UMass Amherst. I build ML that runs under real limits and gets checked against real baselines: a person detector counting dining-hall lines on a CPU at 163 ms a frame, and an injury-risk model on 637K NYC crashes, scored only on crashes it never trained on.

View work Get in touch Open to full-time · Summer 2027
01 Selected Work

Models in production

Sep 2026 Live Real-Time Computer Vision

Hall Check

Counts the people in line at four UMass dining halls from their public camera streams, one frame every two minutes, live since October 26. YOLO11n runs on a CPU and counts people whose feet fall inside a queue polygon. Ultralytics shrinks the 1280×720 streams to 640 by default, which loses the people at the far counters, so I run it at full size. On the same Berkshire frame it found 21 people where 640 found 7, at 163 ms a frame instead of 56 and 522 MiB peak memory instead of 433. The CPU-only PyTorch wheel cut the worker image by more than a third, and CI fails if a CUDA library comes back. Frames are never stored: CI also fails on any pixel-writing call. The accuracy report is fixed in advance and refuses to produce a number until 150+ blind hand counts cover every hall, meal, and lighting condition.

CPU / frame
163 ms
Peak memory
522 MiB
Frames stored
0
  • Python
  • PyTorch
  • Ultralytics YOLO
  • OpenCV
  • Docker
  • Linux
  • Supabase
  • Next.js
Live site ↗ GitHub ↗
2025–2026 Live ML / Calibrated Risk Model

NYC Collisions

A gradient-boosted model estimates how likely each crash was to injure someone from what was involved but not where it happened. On 121,970 held-out 2025–26 crashes it scores ROC-AUC 0.79 against 0.54 for a borough-and-hour base rate. Five rolling-forward fits score 413,780 crashes so none is scored by a model that trained on it. Of 1,093 ranked intersections, 101 clear z = 1.96 where chance predicts about 27, but only 2 survive Benjamini–Hochberg, so the list is a screening queue. A backtest shows the ranking persists, though it only ties a no-model sort. A weekly job rebuilds all 637,256 crashes behind 387 tests.

ROC-AUC
0.79
Brier skill
+26.9%
Sites past z = 1.96
101
  • Python
  • scikit-learn
  • DuckDB
  • Parquet
  • GitHub Actions
  • Streamlit
Live app ↗ Model card ↗ GitHub ↗

Evaluation

Oct 2026 Power Analysis / Multiple Testing

When Evals Can't Tell

How large an LLM eval has to be before a 5-point gap between two prompts means anything. On 100 items, the usual two-proportion test catches a real 80% vs. 85% gap 15% of the time, and 80% power takes 903 items. Both systems answer the same items, so McNemar's paired test applies, and when they disagree on 10% of items it needs 312. Score 50 prompts that are no better than baseline on 300 items, keep the best, and it averages +6.5 points from selection alone. The CLI prints what an eval could detect before any p-value, and CI fails if the regenerated results drift from the committed ones.

Items, unpaired
903
Items, paired
312
Best-of-50 bias
+6.5 pts
  • Python
  • NumPy
  • SciPy
  • statsmodels
  • Power analysis
  • Multiple testing
Results ↗ GitHub ↗
Oct 2026 NLP / Transformers

Text vs. Timing

Does posting time predict engagement beyond what a post says? Three text models run with and without the timestamp on the same split: TF-IDF with logistic regression, frozen MiniLM embeddings under the same linear head, and BERT fine-tuned with a temporal fusion head on a Colab T4. The frozen row separates what pretrained semantics buy from what fine-tuning buys. On 20,000 Stack Overflow posts, no model beats the majority class by more than 1.4 points and fine-tuned BERT ties TF-IDF. Adding the timestamp to TF-IDF fixes 34 predictions and breaks 32 (McNemar p = 0.90), and the generated report calls that inconclusive rather than a gain.

Posts
20,000
Best lift
+1.4 pts
Timestamp p
0.90
  • Python
  • PyTorch
  • Hugging Face Transformers
  • BERT
  • scikit-learn
  • McNemar's test
Results ↗ GitHub ↗
Jun 2026 Experimentation / Causal

When Average Effects Lie

Turns a 64,000-customer randomized email test into a targeting plan. The Women's email's +4.5-point visit lift splits into +7.3 for past buyers of women's merchandise and +1.1 for everyone else, with Benjamini–Hochberg control across the interaction tests. I picked among T-, S-, and DR-learners by 5-fold cross-fitted Qini on training rows. Picking on the reporting split would have quoted a men's-email Qini of 19.3 instead of 2.6. At cost break-even, targeting the Women's email adds $16.95 per 1,000 customers (95% CI $2.99 to $30.60) and sends 415 fewer emails. The Men's email goes to everyone, because its targeting gain can't be told from zero.

Customers
64,000
Gain / 1K
+$16.95
Fewer emails / 1K
415
  • Python
  • scikit-learn
  • statsmodels
  • Uplift modeling
  • Causal inference
  • Bootstrap
Interactive write-up ↗ GitHub ↗

Also built

More projects on GitHub ↗
02 About
Focus
Applied ML and model evaluation
Education
UMass Amherst, Informatics (Data Science) & Psychology
Graduating
May 2027
Based in
Amherst, MA
Projects shipped
15
Records modeled
700K+
Live apps
6
03 Experience
  1. Sep 2026 – Present

    Undergraduate Research Assistant

    Somneuro Lab, UMass Amherst · Amherst, MA

    Onboarded (CITI certified) to the lab's wrist-actigraphy (GGIR) and polysomnography data from studies of sleep and memory consolidation.

  2. Summer 2026 – Present

    Analysis Working Group Member, AI/ML and Brain

    NASA Open Science Data Repository · Remote

    Volunteer member of two of NASA OSDR's open-science Analysis Working Groups: AI/ML, which builds machine learning tools for spaceflight biology, and Brain. In Brain's Behavior & Neuroinflammation subgroup I work on linking a rodent behavioral phenotype from Rodent Research-1 to inflammasome-related neuroinflammation under spaceflight and ground-based radiation. The work pairs OSDR omics and behavioral datasets with environmental telemetry and RadLab radiation records. The subgroup is still defining its research question and analysis plan.

  3. May 2025 – Aug 2025

    AI & Psychology Research Intern

    Spiritual Data · Remote

    Built an automated harness that benchmarked a proprietary model against baselines over repeated runs. Designed evaluation splits and error slices to find failure modes, and wrote the report that informed model selection.

  4. Jun 2022

    Research Intern, Impact of Social Media

    UC Berkeley · Lawrence Hall of Science · Berkeley, CA

    Cleaned and prepared survey data from 700+ participants into analysis-ready datasets, then ran quantitative analysis (EDA, regression) on how social media use relates to public attitudes.

04 Education

University of Massachusetts Amherst

Amherst, MA · Expected May 2027

B.S. Informatics (Data Science concentration) and Psychology

Manning College of Information & Computer Sciences. Coursework: Data Structures, Machine Learning, Databases.

05 Skills

Languages

  • Python
  • SQL
  • Java
  • TypeScript
  • JavaScript
  • R

ML

  • PyTorch
  • Ultralytics YOLO
  • Hugging Face Transformers
  • BERT fine-tuning
  • Sentence embeddings (MiniLM)
  • scikit-learn
  • Gradient boosting
  • Probability calibration
  • Uplift modeling

Evaluation & Statistics

  • Power analysis
  • Paired tests (McNemar)
  • Time-split validation
  • Bootstrap intervals
  • Permutation tests
  • FDR control
  • Empirical Bayes
  • Causal inference
  • A/B testing
  • statsmodels

Data

  • pandas
  • NumPy
  • DuckDB
  • Parquet
  • PostgreSQL

Systems & Tools

  • Linux
  • Docker
  • GitHub Actions (CI/CD)
  • Git
  • pytest
  • FastAPI
  • Streamlit
  • Claude API
06 Contact

Looking for full-time roles from summer 2027.

Machine learning engineering and applied science, especially models that have to fit real latency and memory limits. Email is the fastest way to reach me.

LinkedIn ↗ GitHub ↗ Résumé ↗