Projects

An end-to-end retrieval-augmented generation system that grounds a local LLM in peer-reviewed PubMed abstracts to answer questions on the neurochemistry of mental health, evaluated against the BioASQ-13b biomedical QA benchmark.

BERTScore 0.67 · Answer Correctness 0.60

• RAG
• LlamaIndex
• FAISS
• BioBERT
• llama.cpp
• LLM Evaluation
• RAGAS
• Biomedical NLP
• Local LLMs
• Python

A production-style NLP pipeline that classifies Apple-product feedback by sentiment and product topic, with built-in PII redaction and sarcasm-aware sentiment correction. Served through a FastAPI REST API and a Streamlit UI on a swappable, dependency-injected inference pipeline.

• Python
• PyTorch
• Hugging Face Transformers
• RoBERTa
• FastAPI
• Streamlit
• NLP
• Sentiment Analysis
• PII Redaction
• Sarcasm Detection
• Topic Classification
• REST API

A semi-structured, non-parametric memory framework for RAG that combines knowledge graphs with a hierarchical vector store. My contributions: Query Classification (heuristics → LLM-bootstrapped labels → fine-tuned BioLinkBERT) and Entity-Relation Extraction (REBEL fine-tuned on BioRED, with FastCoref preprocessing) — landed on REBEL after a long experimental funnel.

F1 0.64 · Query Classification on BioASQ

• Python
• PyTorch
• Hugging Face Transformers
• REBEL
• BART
• BioRED
• BioASQ
• BioLinkBERT
• FastCoref
• scispaCy
• Coreference Resolution
• Relation Extraction
• Biomedical NLP
• Knowledge Graphs
• RAG
• Seq2Seq
• Fine-Tuning
• LLM-Based Labeling
• Groq
• NLP

An end-to-end agricultural decision-support project pairing a 22-class crop classifier (Naive Bayes + Random Forest on derived agronomic features) with a soil-moisture data pipeline that computes evapotranspiration from a Penman-Monteith equation, fed by WeatherBit + OpenWeather APIs and persisted to SQLite.

• Python
• pandas
• scikit-learn
• Random Forest
• Naive Bayes
• Classification
• Feature Engineering
• EDA
• Data Visualization
• Flask
• SQLite
• API Integration
• Agronomy
• Penman-Monteith
• Data Science

A systematic study of imbalanced classification techniques on the Home Credit Default Risk dataset (91/9 class split). Final pipeline achieved 0.55 PR-AUC with Logistic Regression and HistGradientBoosting, combining class weighting and threshold tuning. Compared five approaches × three classifiers; Optuna for hyperparameter search; F2 and PR-AUC as the evaluation lens.

• Python
• pandas
• scikit-learn
• Logistic Regression
• HistGradientBoosting
• XGBoost
• imbalanced-learn
• Optuna
• Classification
• Imbalanced Classification
• SMOTE
• ADASYN
• Threshold Tuning
• Class Weighting
• Probability Calibration
• Hyperparameter Tuning
• Feature Scaling
• Feature Hashing
• PCA
• Truncated SVD
• Credit Risk
• Feature Engineering
• EDA
• F2 Score
• PR AUC
• Data Science

A storytelling EDA of the 2022 Indian Premier League season — visualizing toss-decision impact, powerplay performance per team, and batsmen / bowler effectiveness across the three phases of a T20 innings. Built entirely in Plotly (subplots, box plots, scatter quadrants) and published as a Medium article.

• Python
• pandas
• numpy
• Plotly
• Data Visualization
• EDA
• Storytelling
• Sports Analytics
• Cricket
• Jupyter
• Data Science