Hariprasad AP

A CNN model that classifies chest X-rays as Normal or Pneumonia, built from scratch rather than fine-tuned from a pretrained model.

• Designed a custom CNN architecture from scratch, avoiding transfer-learning bias from models trained on natural images.
• Fixed a class imbalance problem (74% vs 26%) using dynamic class weights and orientation-safe augmentation.
• Achieved 90.38% test accuracy, 96.84% recall, and 0.9542 ROC-AUC.
• Built a Streamlit portal so the model's predictions and performance charts are usable, not just numbers in a notebook.

• Python
• TensorFlow
• Keras
• Streamlit
• Scikit-Learn
• NumPy

A fully offline desktop app that indexes files on your machine and lets you search them by meaning, not just filename.

• Generated dense vector embeddings locally using Sentence Transformers.
• Used FAISS for fast similarity search over the embeddings.
• Built the desktop UI in PySide6 — runs entirely on-device, no cloud calls at any point.
• Handles recursive folder traversal, parsing, and indexing with fast query response.
• All embedding inference and storage happen on local hardware, so nothing you search ever leaves your machine.

• Python
• PySide6
• Sentence Transformers
• FAISS
• SQLite

A web scraper you talk to in plain language — it figures out what to extract and returns structured data.

• Built the scraping layer with Selenium and BeautifulSoup to handle dynamic, JS-rendered pages.
• Used the Gemini API to turn a natural-language request into structured, schema-based extraction.
• Built a Streamlit interface with automated DOM chunking and content cleaning.
• Benchmarked a local Ollama/DeepSeek setup before switching to Gemini API for better latency — kept the local option in mind for cost and privacy tradeoffs.

• Python
• Selenium
• BeautifulSoup
• Streamlit
• Gemini API

A university project exploring how blockchain concepts could secure a digital voting system. I built the backend.

• Built the backend with FastAPI and MySQL to handle user data and voting logic.
• Implemented a basic blockchain structure to record votes as tamper-evident entries.
• Frontend was HTML/CSS/JS with a Flutter-based responsive layer, built alongside teammates.

• Python
• FastAPI
• MySQL
• Flutter
• HTML/CSS/JS

A public learning diary where I write out CS fundamentals and DSA concepts in my own words, to check if I actually understand them.

• Built with MkDocs Material, hosted on GitHub Pages.
• Every concept is explained the way I'd explain it to someone else — the Feynman technique, used as a real check on my own understanding.
• Includes build logs from my other projects — what broke, what I changed, and why.
• Global search, clean code blocks, and a reading experience I actually wanted to use myself.

• MkDocs Material
• Markdown
• Git
• GitHub Pages

A simple desktop flashcard app using spaced repetition, built because I wanted one that worked the way I studied.

• Built the GUI in Tkinter for reviewing custom vocabulary sets.
• Used pandas and CSV files to track and log review progress over time.
• Added timed card flips and spaced review intervals.

• Python
• Tkinter
• Pandas
• CSV Automation