I build data systems end to end, from a 25 GB Spark pipeline to the model and the dashboard someone actually opens. Before grad school I spent a year shipping production software: a hospital information system used by 1,000+ healthcare workers, and a sales-analytics platform where I went from intern to leading a team of five.
What I care about: picking the metric that matters, not the one that's easy. It keeps coming up in my work: ranked by clinical cost, a 5th-place heart-disease model saves the most lives, and in my trading backtests the two least accurate models made the most money.
- 🎓 M.S. Applied Data Analytics, Boston University (GPA 3.8)
- 🔭 Looking for Data Science / ML / Data Engineering internships and new-grad roles
- 📫 LinkedIn · kunjpatel271002@gmail.com · Boston, MA
WikiFlow: Predicting Wikipedia editor dropout from 60M revision events
Logistic regression and K-Means written from scratch on Spark RDDs, a distributed Keras DNN, and live scoring of the Wikipedia edit stream through Kafka. Runs on GCP Dataproc and BigQuery. PySpark Kafka GCP Keras
Beyond Accuracy: Choosing a heart-disease model by what mistakes actually cost
Six classifiers on 70K patients. Once they're ranked by clinical cost, a model that placed 5th on accuracy saves $9.7M and 65 missed diagnoses per 10K patients. Includes fairness, robustness and sensitivity audits. scikit-learn pandas
RAG Document Q&A: Ask questions across PDFs, web pages, Wikipedia and arXiv
LangChain pipeline with pluggable vector stores (Chroma, FAISS, ObjectBox) and LLMs (OpenAI, Groq, local), served through a Streamlit UI and a FastAPI/LangServe API. LangChain FastAPI Streamlit
Cassandra vs MongoDB: Benchmarking four workloads on 13M NYC taxi trips
Cassandra writes 2.2× faster; MongoDB deletes 30× faster and aggregates 5× faster. Also covers the async-write mistake that made my first results wrong. Cassandra MongoDB GCP
Football Market Value: What actually drives a player's transfer price?
Regression, ANOVA, chi-square and odds ratios on 900 top-5-league players. Finds an "aging star discount" of about €14M. R
| Project | Result | |
|---|---|---|
| 📈 | Can ML beat buy-and-hold? | Backtested 10 classifiers on 5 years of Coca-Cola stock; only 3 beat buy-and-hold, and the least accurate earned the most |
| ⚡ | Big data algorithms in PySpark | External merge sort, NYC-taxi analytics, TF-IDF kNN and gradient descent written on RDDs, without MLlib |
| 🏠 | Disability prediction, AHS 2023 | 70+ models on Census data; undersampling beat SMOTE, and gradient boosting catches 73.6% of disability households (AUC 0.84) (team project) |
| 🫀 | Fetal health from CTG | 88.8% accuracy, but only 63% of abnormal cases caught: why imbalance hides failure |
| 🏢 | BU building energy benchmarks | 82 BU buildings run about 40% above Boston's average energy intensity (city BERDO data) |
| 💵 | Banknote forgery detection | From a 71% hand-written rule to 100% with kNN, plus a feature-ablation study |
| 🔗 | Clustering Dow Jones stocks | Residual-return k-means over 60 months; sector peers don't reliably move together |
| 🧰 | Data mining in R | Feature selection (CFS, Boruta, RFE), LASSO, caret ensembles, Apriori, k-means |
Software Developer · Artem HealthTech 2024
Cut page-load time 35% on a hospital information system used by 1,000+ healthcare workers. A Redis caching layer reduced database load 40% and saved $3K/month.
Project Leader · Wellnest Tech 2024 · promoted from intern in 3 months
Led 5 developers on an Angular/.NET sales-analytics platform for 200+ daily users. Built CI/CD that cut deploys from 3 days to 4 hours.
Software Developer Intern · Telnet 2023
Built an Angular recipe platform for 500+ beta users. Its one-click cart reached 70% adoption in two weeks.
Daily Python · SQL · pandas · scikit-learn · PySpark · Git · Jupyter
Also shipped with R · Kafka · LangChain · FastAPI · Streamlit · TensorFlow/Keras · XGBoost · FAISS
Data & cloud PostgreSQL · MySQL · MongoDB · Cassandra · Redis · BigQuery · GCP · AWS (S3, EC2) · Docker
BI & web Tableau · Power BI · Angular · .NET · TypeScript