Four systems I've built from scratch — retrieval, tabular ML, speech and vision.
Finance RAG Pipeline
LangChain · ChromaDB · Python
A retrieval-augmented question answering pipeline over financial documents: chunking, embedding-based retrieval from a vector store, and an LLM answer stage. Built as the first piece of a larger multi-agent orchestration system.
View repository →
Health Data Classification — NHANES
XGBoost · LightGBM · CatBoost · Optuna · SHAP
A binary classification pipeline on NHANES health survey data, built during the Summer Analytics 2026 Hackathon (IITG Consulting & Analytics Club). Class imbalance handled with SMOTE, tuning with Optuna, three boosters combined in a soft-voting ensemble, and SHAP used to explain what the model actually keys on. Now being extended into a comparative research paper on explainable ML.
Voice Intent Classifier
Whisper · BERT · PyTorch
Speech in, intent out. Whisper transcribes the audio, a fine-tuned BERT classifier maps the transcript onto intent labels — the front half of any voice-driven assistant.
CNN Image Classifier
PyTorch · CIFAR-10 · Flask
A convolutional network trained on CIFAR-10 on a Colab T4 GPU, then wrapped in a Flask interface so predictions run in a browser rather than a notebook.