Every project I take on starts with a simple question: “What problem am I really solving?” Below are some of my
favorite projects, explained through the challenges they addressed, the routes I took, and the impact they delivered.
🧾 RAG-Based Chatbot for Life Insurance Policy Understanding
Insurance policies are notoriously difficult to read and are full of jargon, dense tables, and scattered clauses. I
wanted to make this process human-friendly.
The Route I Took:
- Extracted text and tables from policy PDFs using pdfplumber.
- Broke them into 500-token chunks with 50-token overlap for better contextual retrieval.
- Stored embeddings with OpenAI ada-002 vectors inside ChromaDB.
- Designed a dual-stage retrieval system: first vector similarity, then a CrossEncoder MiniLM reranker for accuracy.
- Finally, passed refined results into GPT-3.5 Turbo, with hallucination controls to ensure answers stayed grounded.
The Result: A chatbot that can answer real-world queries like “Can I surrender this policy?” or “How is the
claim processed?” with clarity and accuracy. It transformed frustrating policy documents into an interactive Q&A tool
making insurance more accessible.
🚲 Bike Demand Prediction
For companies like Boom Bikes, the challenge is not just selling bikes it’s predicting how many will be rented
tomorrow. Overstocking is wasteful, understocking is bad business.
The Route I Took:
- Collected weather, seasonal, and user data.
- Used Scikit-learn with a Multiple Linear Regression (MLR) model.
- Applied VIF-based feature pruning and scaling to remove multicollinearity.
The Result: An R² score of 0.82 strong predictive power. This meant automated inventory management became
feasible, helping optimize fleet availability and reducing lost revenue.
🩺 Skin Cancer Detection with CNN
Dermatologists face the challenge of distinguishing between multiple skin cancer subtypes early on where accurate
detection matters most.
The Route I Took:
- Built a TensorFlow + Keras CNN pipeline.
- Applied rotation and scaling augmentations to enrich training data.
- Added dropout layers to reduce overfitting.
The Result: Achieved 92% test accuracy across 7 skin cancer subtypes. This project demonstrated how AI
could support doctors in early diagnosis, potentially saving lives.
✋ Gesture Recognition with 3D-CNN
Human-computer interaction is evolving and I wanted to explore how machines could see and understand gestures.
The Route I Took:
- Designed a Conv3D + ConvLSTM spatiotemporal model.
- Tuned hyperparameters like learning rate and batch size.
- Regularized with L2 penalties to stabilize training.
The Result: Achieved 78% validation accuracy on a challenging video dataset. The project laid the foundation
for gesture-controlled interfaces in AR/VR environments.
🌬️ Cylindrical Filter Simulator – Air Filtration System Modeling Tool
Industrial filters are everywhere — but optimizing them is still largely trial and error. I wanted to bring
science + AI into filtration design.
The Route I Took:
- Built a multi-physics simulator in Python.
- Modeled airflow, pressure drop, dust loading, and efficiency.
- Added geometry optimization features and a real-world media database.
- Exported publication-ready plots (Matplotlib), reports (ReportLab PDF), and datasets (JSON/CSV).
The Result: A research-grade tool with Colab UI integration, making it accessible to scientists and
engineers. It showcased how AI-powered design tools can accelerate innovation in manufacturing.
✈️ Travel Experience Chatbot (Ongoing)
Planning a trip means juggling multiple apps: flights, hotels, weather, itineraries. I asked: “What if one AI could
do it all?”
The Route I Took:
- Engineered a Flask-based chatbot.
- Integrated OpenAI GPT, Amadeus API, and weather APIs.
- Applied prompt engineering for personalized itineraries.
The Result (so far): A single interface that gives users real-time, customized travel plans a project I’m still
refining, but one that excites me for its everyday usefulness.
📰 Semantic Fake News Detection
The spread of fake news is a global challenge. I wanted to see if semantics the meaning of words could help
machines separate fact from fiction.
The Route I Took:
- Preprocessed text with lemmatization + POS tagging.
- Generated embeddings using Word2Vec.
- Trained classifiers: Logistic Regression, Random Forest, and Decision Tree.
The Result: The Logistic Regression pipeline reached 91% accuracy and 0.906 F1 score. The project proved
how simple yet semantic-rich models can fight misinformation effectively.
✨ Each of these projects started as curiosity but ended up becoming solutions that others could use, learn from, or build upon. My hope is that they inspire you to imagine your own because every project starts with a single “what if?”.