Hey, I'm Maya.

I like taking messy, real-world datasets and finding the story inside them — through exploratory analysis, predictive models, and the occasional applied AI experiment. Take a look at a few fun projects of mine:

90K+
records modeled
0.938
R² on salary regression
172K
rows cleaned & analyzed
3
deployed / published projects

From raw tables to real decisions

Every project below follows the same arc: understand the data, clean it honestly, engineer what matters, then build something you can actually interact with.

Data Wrangling & EDA

Cleaning, imputing, and interrogating data until it tells a trustworthy story — with visuals that make the finding obvious.

Machine Learning & Prediction

Feature engineering, clustering, and model tuning — regression and classification pipelines built and evaluated end-to-end.

Applied AI

Using embeddings and pretrained models (like CLIP) to build multimodal tools, then shipping them as real, usable apps.

Projects

Three projects, three different muscles: applied ML, multimodal AI, and analytical storytelling.

01 / MACHINE LEARNING

Global AI & Data Salary Prediction

Pythonscikit-learnRandom Forest K-MeansFeature EngineeringGradio

An end-to-end pipeline predicting salaries across ~90,000 AI & data roles worldwide (2020–2026). Cleaned outliers with the IQR method, engineered K-Means "career archetype" clusters and polynomial features, then trained and tuned Random Forest models for both salary regression and a 3-tier salary classifier — deployed as a live interactive app.

0.938
R² score
$7,999
MAE
89%
classification accuracy
02 / APPLIED AI · MULTIMODAL

Find Your Michi 道

CLIPEmbeddingsK-Means Cosine SimilarityGradio

A multimodal career-matching tool: describe or photograph your ideal workspace, and it's converted into a CLIP embedding and compared against 1,500 indoor-scene images mapped to O*NET's 923 occupations via the RIASEC personality framework. Clustered the embedding space (K=5) and built a real-time cosine-similarity recommendation engine.

1,500
image embeddings
923
O*NET occupations
768-d
CLIP vector space
03 / EXPLORATORY DATA ANALYSIS

Horse Race Prediction — EDA

PandasData CleaningStatistical AnalysisData Storytelling

A full exploratory analysis of 172,000 historical race records to find what actually predicts a win. Handled missing data with targeted imputation, engineered an "actual odds" feature from raw betting probabilities, and used correlation analysis to isolate the strongest signals — distilled into a clear, evidence-backed profile of a winning horse.

172K
race records
0.42
top correlation (RPR)
55.7%
favorite win rate

Skills & Tools

LANGUAGES & CORE TOOLS

Python Pandas NumPy Jupyter Git SQL

MACHINE LEARNING

scikit-learn Random Forest K-Means / DBSCAN Feature Engineering GridSearchCV

DATA ANALYSIS & VIZ

Exploratory Data Analysis Matplotlib / Seaborn Statistical Analysis Data Storytelling

APPLIED AI

Hugging Face CLIP / Embeddings Gradio Model Deployment