Projects

Selected engineering and applied machine learning projects spanning retrieval systems, geospatial modeling, and full-stack AI deployment.

Kalinga

Multilingual Hospital Request Routing

Rutgers HealthHack 2026 — Honorable MentionPythonPyTorchHugging FaceSetFitscikit-learn
  • Developed a multilingual, multi-label NLP classifier for hospital request routing across 4 languages and 6 departments; curated a labeled dataset and benchmarked TF-IDF/logistic regression and BGE embeddings.
  • Fine-tuned multilingual MiniLM using SetFit and contrastive learning; achieved 0.8963 macro F1 / 83.73% accuracy on grouped test data and 0.8339 macro F1 / 79.03% accuracy on an independent holdout set, using StratifiedGroupKFold for group-aware splitting.
  • Implemented emergency escalation and out-of-scope abstention rules for safety-aware routing; published the fine-tuned classifier on Hugging Face.

Knowledge Distillation for Reasoning in MoE and Dense LLMs

Transferring chain-of-thought reasoning from large LLMs to smaller models

PythonPyTorchHugging FaceLoRA
  • Developed sequence-based and logit-based knowledge distillation pipelines to transfer chain-of-thought reasoning from GPT-OSS 120B and LLaMA 70B teacher models to a 3B student using LoRA-based fine-tuning.
  • Implemented temperature-scaled KL divergence, top-k logit extraction, and cross-tokenizer alignment to match teacher-student output distributions despite vocabulary mismatches.
  • Achieved up to 4.5% accuracy improvement and 96%+ BERTScore on SciBench and TheoremQA; demonstrated that logit distillation consistently outperformed sequence-based distillation.

GenreBlender

Generative AI Music Mixer & Classifier

PythonPyTorchMusicGenStreamlitscikit-learn
  • Architected an end-to-end generative AI system using Meta’s MusicGen to synthesize 10-second hybrid audio clips with controllable genre blending via a continuous weighting parameter (α ∈ [0,1]).
  • Engineered a GTZAN preprocessing pipeline with feature selection, normalization, and GroupShuffleSplit validation; trained a custom 4-layer PyTorch MLP achieving 92% validation accuracy.
  • Developed a probability-based evaluation framework comparing predicted genre distributions against weighted target mixtures; deployed an interactive Streamlit application for audio generation and quantitative evaluation.

PocketRAG

Retrieval-Augmented Generation System

  • Built a lightweight RAG web application enabling instant PDF summarization and document-based question answering.
  • Implemented semantic search using FAISS vector indexing with embedding-based chunk retrieval.
  • Integrated Gemini API for context-aware generation with dynamic prompt construction.
  • Designed clean frontend UI for real-time query-response streaming and document upload.

Telematics UBI Prototype

Usage-Based Insurance Modeling Platform

  • Developed a full-stack usage-based insurance system for telematics risk scoring and driver behavior analysis.
  • Built XGBoost-based risk prediction models trained on telematics driving data with engineered behavioral features.
  • Implemented FastAPI backend for model serving and REST endpoints with secure AWS deployment.
  • Designed React frontend dashboard visualizing driver risk scores and prediction outputs.

This Website!

Personal Portfolio

  • Designed and deployed a responsive personal portfolio website using Next.js, Tailwind CSS, and TypeScript to showcase my work in AI/ML.
  • Implemented GitHub-based CI/CD with Vercel for automated build and deployment on every commit.
  • Leveraged modern React architecture (App Router) and optimized performance using built-in Next.js font and asset optimization features.