Arch Desai Portfolio
TAMU Datathon 2020 winner (among 50+ teams) with a PHM Society publication. Structured GitHub portfolio covering NLP, time-series, survival analysis and hand-built statistical methods — a strong template for new data scientists.
Location
Unlisted — probably wherever the wifi is strongest.
Skills
What makes it work
A breakdown of the choices that make this portfolio stand out.
Competition wins as hard credibility anchors
A datathon win is a public, verifiable, competitive result. Unlike self-reported project descriptions, competition placements are adjudicated by independent judges against many other candidates simultaneously. They are objective proof of performance under pressure.
Topic breadth with clear structure
The portfolio covers NLP, time-series analysis, and classical ML — not by listing keywords, but by including actual projects in each area. The structure maps directly to the types of problems companies actually have.
Problem → approach → result framing
Each project is presented as: what was the business problem, what data/method was used, what metric improved. This framing is exactly what data science interviews probe — it signals that the candidate understands that models serve business outcomes.
What Data Scientists can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
Enter Kaggle competitions or datathons. A placement is public proof of skill that no self-built project can replicate.
Structure every project the same way: business problem, data source, methodology choice (and why), metric result. Consistency makes scanning easy.
Cover NLP, time-series, and tabular ML in your portfolio. These three areas cover the majority of real industry problems.
Don't just show accuracy scores. Show how the metric improvement translates to business value.
A GitHub profile with 12 well-documented projects beats 40 repositories that are just notebooks with no READMEs.
Projects
Beneath the eight headline projects sits the section most portfolios skip. Arch Desai’s micro projects implement gradient descent, expectation maximization and a neural network by hand, then rank NFL teams with Markov chains — the results up top, the math underneath.
Telecom customer churn studied two ways. Survival analysis (Kaplan-Meier curves, log-rank tests by customer group, a Cox proportional hazards model) shows how churn likelihood changes over time. A Random Forest then predicts churn for individual customers. A Flask app shows a customer's churn gauge, hazard and survival curves, SHAP explanation and expected lifetime value.
Pairing "will they churn?" with "when, and what are they worth?" is the framing retention teams actually use. The SHAP view per customer shows explainability in the product, not just in the notebook.
Survival curves and hazard model alongside a classifier, not instead of one
Per-customer SHAP explanation and lifetime-value estimate in the app
Repo organized with separate EDA, survival and model notebooks plus app code
Engine failure and remaining-useful-life prediction on NASA C-MAPSS turbofan data: 100 training engines with runs of 128–356 cycles and 100 separate test engines. The project covers seven approaches: an exponential degradation model and a similarity-based model for RUL, an LSTM for RUL, LSTM / RNN / 1D-CNN for binary and multiclass failure classification, and a 1D CNN-SVM.
Comparing classical degradation models with deep sequence models on the same fleet data demonstrates time-series judgment. This is directly relevant to industrial and IoT data-science roles.
Frames the problem both as classification (fail within n cycles) and as regression (RUL)
Seven models, from exponential degradation to CNN-SVM
Train and test engines are distinct, so there is no leakage across units
A personalized news recommender that combines collaborative filtering (Twitter users with similar interests) with content-based filtering (similarity between articles and a user's tweets). The flow: collect active news readers who retweet the NYT, Bloomberg and the Washington Post, filtered by a follower/friend popularity index. Then clean and lemmatize their tweets, cluster users on TF-IDF vectors, run sentiment analysis and topic modeling, and match articles to a given Twitter handle.
Hybrid filtering built from raw social data shows NLP pre-processing, unsupervised clustering and recommender design in one project, and the data is collected rather than downloaded.
Collaborative and content-based signals combined
Custom popularity index used to filter real news readers
Step-by-step README walks through each pipeline stage
Ranking NFL Teams with Markov-Chain Methods
One of the "micro projects" beneath the headline work. It implements and compares three Markov-chain stationary-distribution approaches that rank all 32 NFL teams from best to worst using scores from the 2007 regular season, written up as a report.
Implementing ranking math from first principles, not calling a library, shows the mathematical footing interviewers probe for in statistics-heavy roles.
Three stationary-distribution methods compared on the same season
Ranking from game scores alone — no pre-built ratings
Part of a micro-projects section of from-scratch algorithm implementations
More Data Scientist portfolios
Ready to build your portfolio?
Follow the Data Scientist roadmap — skills, projects, and timeline to get hired.