Ritesh Penumatsa

• Generic resumes get filtered out before a human ever reads them.
• Solo — design and ship a tool that closes that gap.
• Built a full-stack app with a serverless proxy and a dual PDF-parsing pipeline with an AI fallback.
• Shipped to production at resumematch.org — learned to design for reliability at scale, not just a working demo.

Tools: Claude API · serverless functions · PDF parsing

• Austin traffic patterns and their drivers weren't well understood from the raw data.
• One of 5 teammates, leading the modeling and interpretability work.
• Ran hypothesis tests, built Linear Regression and Random Forest models, applied SHAP.
• Weather stood out as a key driver — learned to explain "black-box" model decisions with SHAP.

Tools: Python · scikit-learn · SHAP · Jupyter

• Unclear whether bed-net (ITN) distribution actually lowers child mortality.
• One of 3 teammates, co-leading modeling and fairness evaluation.
• Built a Random Forest model, checked subgroup fairness with Brier scores, built the dashboard.
• Found only a weak link (r = 0.13, p = 0.32) — learned correlation isn't causation, and to check fairness across subgroups.

Tools: Python · Random Forest · Brier score · dashboarding

• Misinformation spreads faster than people can fact-check it.
• Solo — build and ship a working fake-news classifier.
• Built an NLP pipeline with a Gradient Boosting Classifier, deployed via Streamlit.
• 98% accuracy, 35% fewer false positives — learned text preprocessing and shipping a model end-to-end.

Tools: Python · scikit-learn · Streamlit

• The team wanted to know what actually drives ChatGPT's app ratings.
• Solo — turn raw review text into a predictive model.
• Scored sentiment, engineered features from review text, trained a regression model.
• 99%+ accuracy — learned feature engineering can matter more than the model itself.

Tools: Python · scikit-learn · Matplotlib/Seaborn

• Farmers need a fast way to catch tomato plant disease early.
• Solo — build and deploy an image classifier people could actually use.
• Trained a CNN on labeled leaf images, built a Streamlit web app around it.
• 95% accuracy, ~50% faster diagnosis — learned computer vision deployment end-to-end.

Tools: Python · CNN · Google Colab · Streamlit

• People want a quick, private way to gauge their diabetes risk.
• Solo — turn a health dataset into a self-assessment tool.
• Trained and tuned a classifier on age, BMI, and glucose, wrapped in a Streamlit interface.
• 95%+ accuracy, used by 100+ people — learned to balance model reliability with an approachable UI.

Tools: Python · scikit-learn · Streamlit

• Wanted to test how well classical feature-based CV could identify people, not just deep learning.
• Solo — build a working classifier and a demo for it.
• Extracted facial features with OpenCV, trained an SVM, built a Flask app.
• 92% accuracy — learned the trade-offs between classical feature engineering and deep learning.

Tools: Python · OpenCV · scikit-learn (SVM) · Flask

• Execs needed faster, data-backed answers to ad hoc business questions.
• Solo — simulate a junior analyst role: answer 10 SQL prompts.
• Wrote and optimized queries for sales, inventory, and customer segmentation.
• Delivered 10+ documented insights — learned to write SQL for a business audience, not just correctness.

Tools: MySQL Workbench · Google Sheets/Slides