Projects

A selection of things I've built, from medical imaging to LLM applications

Liver Tumor Segmentation Deep Learning Pipeline
Liver Tumor Segmentation Deep Learning Pipeline

Built a parallelized U-Net training pipeline for liver tumor segmentation on 53GB of CT scan data, comparing data- and model-parallel training across multiple GPUs and cutting training time by up to 29% while maintaining ~98% validation accuracy, with MLflow and DVC for experiment tracking.

U-NetMLflowDagsHubAWS EC2
Product Recommendation Engine
Product Recommendation Engine

Built a GPT-4-powered recommendation engine that generates personalized product suggestions from user preferences and browsing behavior, deployed with a Flask/React stack.

FlaskReactHerokuVercelchat-GPT4o
Depth Vision MLOps Pipeline
Depth Vision MLOps Pipeline

Built a full-stack computer vision app combining Depth-Anything-V2 monocular depth estimation and YOLOv5 object detection as containerized microservices, deployed to AWS ECS behind an autoscaling load balancer and provisioned entirely with Terraform and CodeBuild CI/CD.

Depth-Anything-V2YOLOv5TerraformAWS ECSDocker
Molecular Property Prediction
Molecular Property Prediction

Built graph neural networks (GCN, GAT) to predict drug permeability from molecular structure, reaching 85%+ accuracy and 0.85 ROC-AUC.

GCNGATPyTorch Geometric
Bidirectional MRI Synthesis
Bidirectional MRI Synthesis

Trained a timestep-aware diffusion model (DDPM) with a cross-attention U-Net to synthesize MRI T1 to T2 modalities, reaching PSNR 23.8 / SSIM 0.81.

UnetDDPM
Conditional SMILES Generation Using GPT2
Conditional SMILES Generation Using GPT2

Fine-tuned GPT-2 to generate novel drug-like molecules (SMILES strings) conditioned on target biological properties, with parallelized descriptor calculation for a 20x speedup.

GPT2Parallel Computation
Chest X-ray Report Generation
Chest X-ray Report Generation

Fine-tuned a ViT-GPT2 hybrid model to generate free-text radiology reports directly from chest X-ray images.

VITGPT-2
Conversational AI with RLHF
Conversational AI with RLHF

Fine-tuned GPT-2 with a full RLHF pipeline, training a Bradley-Terry reward model on preference pairs and optimizing the policy with PPO, achieving a 1,065% improvement in model alignment scores.

GPT-2PPORLHFPyTorch
Generative Design Module
Generative Design Module

Built a generative CAD design pipeline using VAEs to learn surface and edge representations of B-rep parts, paired with latent diffusion models to generate new CAD-compatible mechanical components satisfying physical constraints.

VAELatent DiffusionCADPyTorch
Agentic HR Assistant
Agentic HR Assistant

Built a LangGraph-based HR assistant that reasons step-by-step over multi-turn conversations, using persistent context memory to automate hiring workflows.

LangGraphGPT-4oStreamlit
Research Assistant
Research Assistant

Built an end-to-end RAG web app powered by Mistral-7B, combining BM25 lexical search with FAISS dense retrieval (MiniLM-L6-v2 embeddings) for grounded question answering.

Mistral-7BFAISSBM25AWS EC2
Adaptive Image Processing
Adaptive Image Processing

Built an adaptive image-restoration pipeline combining CLAHE and noise-aware filtering, recovering 95% of image quality across varying noise levels and doubling signal-to-noise ratio.

GoogleNetCLAHEImage Filtering
Text Generator
Text Generator

Built an RNN-based text generation model for Kannada using custom-trained CBOW embeddings for richer word representation.

RNNCBOW EmbeddingTensorFlow