Projects

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

Knowledge Distillation for Reasoning in MoE and Dense LLMs

Efficiently transferring reasoning from large LLMs to smaller models using parameter-efficient distillation

Developed an offline knowledge distillation framework combining sequence-level and logit-level distillation to transfer reasoning capabilities from large teacher models to smaller student models.
Leveraged teacher-generated chain-of-thought (CoT) reasoning traces and trained student models via supervised fine-tuning to replicate both intermediate reasoning steps and final answers.
Applied Low-Rank Adaptation (LoRA) to enable parameter-efficient distillation, updating less than 1% of model weights while preserving performance and reducing compute requirements.
Implemented logit-based distillation using temperature-scaled softmax and KL divergence loss to align student token distributions with teacher outputs via top-k probability reconstruction.
Built an offline pipeline to extract teacher logits and map them to student token spaces, enabling scalable training without direct API dependency.
Evaluated performance on SciBench and TheoremQA across multi-domain reasoning tasks, using accuracy and BERTScore to measure both correctness and reasoning alignment.
Demonstrated that small LLMs can approximate both the reasoning process and output distributions of larger models with minimal parameter updates.

Generative AI Music Mixer & Classifier

Built a generative AI system that blends two music genres using a controllable slider (α ∈ [0,1]) and generates hybrid audio with Meta’s MusicGen.
Designed a clean ML pipeline on the GTZAN dataset, including feature selection, scaling, label encoding, and reproducible preprocessing for inference.
Trained a 4-layer PyTorch MLP (92% validation accuracy) using scikit-learn’s GroupShuffleSplit to prevent data leakage and predict genre probabilities from audio features.
Developed a measurable genre-blending framework by comparing predicted probabilities against a weighted target distribution to quantify how well the blend worked via Target = (α·A) + (1−α)·B.
Built and deployed an interactive Streamlit web application enabling real-time genre blending, probabilistic evaluation, and 10-second AI-generated audio output.

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.

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.

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