Natassha Selvaraj Portfolio
Data scientist portfolio that prioritises human-interest project selection — gender disparity in Hollywood, celebrity face recognition, heart disease risk factors — making technical work immediately compelling to non-technical hiring managers.
Location
Location: classified. Send a pigeon.
Skills
What makes it work
A breakdown of the choices that make this portfolio stand out.
Human-interest project framing
Projects like "Gender Disparity in Hollywood" and "Heart Disease Risk Factors" are chosen for immediate emotional resonance, not just technical impressiveness. A hiring manager who cannot evaluate a confusion matrix can immediately understand why the question matters — and that accessibility is a huge advantage in mixed-audience portfolio reviews.
End-to-end pipeline on every project
Several projects run end to end: data gathered from multiple sources (APIs, CSVs, and in one case a custom web scraper), then EDA, a model, and a front end — Flask for the celebrity app, Dash for the Harry Potter classifier, a GitHub Pages site and a Tableau dashboard for the movie project. The full pipeline, not just the model, is the proof of data engineering capability alongside data science capability.
Public writing as a multiplier
Natassha's published write-ups for each project (in Towards Data Science, The Startup, Better Programming and Data Driven Investor) create discoverability that the portfolio site alone cannot generate. SEO-indexed explanations of her work surface in Google searches for data science project ideas, bringing inbound traffic from exactly the people most likely to hire or collaborate with her.
Breadth of technique across projects
NLP, computer vision, web scraping, EDA, classification, and recommender work appear across the portfolio without feeling scattered. The connecting thread is storytelling through data — every technique is in service of answering a human question, which gives the breadth coherence.
What Data Scientists can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
Choose projects based on how compellingly you can explain the WHY to a non-technical person. The best data science portfolios work in an interview with a CFO and an ML engineer simultaneously.
Build a deployed interactive application for at least two projects — a Streamlit app, a Dash dashboard, or a web frontend. Showing live output is categorically more impressive than a Jupyter notebook in a GitHub repo.
Write about your projects on Medium or Towards Data Science. A published article creates compounding discoverability and forces you to explain your work clearly, which is the skill gap most data science candidates have.
Include web scraping or data engineering work in at least one project. Employers know data is rarely clean and pre-packaged — showing that you can source and prepare data from scratch proves readiness for real-world work.
Structure each project README with three sections: the question you asked, the approach you took, and what you found. The question is what makes people want to read the rest.
Projects
Each project opens on a question a stranger would want answered — who’s your celebrity lookalike, can you spot an AI-generated face — and ends with the write-up. Natassha Selvaraj treats the article as part of the build, which is how work reaches readers who never open a notebook.
Celebrity Facial Recognition App
"Have you ever wondered who your celebrity lookalike was?" Users upload a photo and click predict. A Flask endpoint decodes the image, detects the face with MTCNN, crops and resizes it to 224×224, and passes it to a pretrained VGGFace ResNet-50 model that predicts the closest-matching celebrity identity. The front end is an HTML/Bootstrap/jQuery page.
A working detect → crop → predict pipeline behind an API endpoint, with a front end anyone can use, shows the serving side of deep learning that notebook-only projects skip.
Face detection (MTCNN) before classification, not raw whole-image input
Pretrained VGGFace model reused rather than trained from scratch — a sensible scoping call
Hook question ("who is your celebrity lookalike?") anyone understands instantly
The Bechdel Test: Analyzing Gender Disparity in Hollywood
Female representation in Hollywood, analysed through the Bechdel Test using data from multiple sources. Bechdel scores come from the bechdeltest.com API (films from 1967 on) and are merged with IMDb ratings and a second movie dataset with director, budget, genre and gross. Director gender is inferred from first names with gender-guesser, and Bechdel pass rates are plotted against year, rating, budget, genre and box-office gross.
Joining an API source with two CSVs, and deriving a new variable (director gender) to answer a question, is the wrangling most analyst work actually involves. The topic needs no domain expertise to follow.
Three sources merged: bechdeltest.com API, IMDb ratings, and a budget/gross/director dataset
Derived feature: director gender inferred from first names
Pass rates compared against budget, gross, genre and IMDb rating
Harry Potter Personality Prediction
A fastai text-classification model trained on an MBTI personality-prediction dataset. Its predictions drive a Harry Potter character classification model, served through a Dash front-end interface.
Taking a standard text-classification task and wrapping it in a playful, instantly understandable theme shows presentation instinct alongside the NLP work.
Text classification with fastai on MBTI-labelled data
Predictions mapped to Harry Potter characters for a memorable hook
Model exposed through a Dash front end, not left in a notebook
End-to-End Movie Project: Dashboard + Recommender
Movie data compiled from several sources into one dataset, then turned into three pieces. First, a Tableau dashboard filterable by reviews, revenue, duration, release year, genre and Bechdel Test result. Second, a recommender that returns five similar movies for a typed title. Third, a project web page tying the dashboard and recommender together.
Data collection, a BI dashboard, an ML recommender and a front page in one project covers the full analyst-to-data-scientist arc, and part of it is still clickable.
Public Tableau dashboard with filters on six attributes
Reuses her Bechdel Test data as a dashboard filter, connecting projects
Dashboard and recommender designed to be used together: find a title, then get five similar ones
More Data Scientist portfolios
Ready to build your portfolio?
Follow the Data Scientist roadmap — skills, projects, and timeline to get hired.