Building damage classification from pre/post disaster satellite imagery

Neural network and backpropagation implemented in pure NumPy with visual explanations

RoPE, ALiBi, sinusoidal, and learned positional encodings — compared experimentally

Clean implementations of 20 landmark CV papers: ResNet through DiT, each in its own module

Attention map visualization across ViT, DINO, and medical imaging models

Focal loss, Dice loss, Tversky, contrastive losses — implemented and visually compared

DDPM with step-by-step mathematical walkthrough as interactive Jupyter notebooks

Micrograd-style automatic differentiation engine built from scratch in Python

Vision Transformer (ViT) trained from scratch on CIFAR-100 with architecture ablations

Full transformer implementation with attention, multi-head attention, encoder-decoder, annotated

Feature store implementation for image embeddings using Redis and Feast

DVC pipeline for large image dataset versioning, lineage tracking, and reproducibility

GitHub Actions CI/CD pipeline for ML: train → evaluate → deploy with quality gates

Custom experiment tracking platform built on MLflow with Optuna hyperparameter optimization

FSDP and DeepSpeed ZeRO distributed training template for large vision models

NVIDIA Triton inference server setup with dynamic batching and model ensembles

Production model monitoring with PSI/KL drift detection and prediction drift alerts

Opinionated ML project template: MLflow + Hydra + PyTorch Lightning + DVC

Predicting cellular response to genetic and chemical perturbations from microscopy

Automated high-content screening pipeline for phenotypic drug discovery

GNN-based drug-target interaction prediction with binding affinity estimation

Graph Neural Network for molecular property prediction on QM9 and ZINC datasets

High-content screening image analysis for cell morphology and viability

Handwritten text recognition with TrOCR fine-tuned on domain-specific handwriting

Automated invoice parsing and structured data extraction using Donut

Table detection and structured extraction from documents and PDFs

Production OCR pipeline using TrOCR and PaddleOCR with post-processing

LayoutLMv3 for invoice, contract, and form parsing with key-value extraction

Self-supervised learning for low-annotation surface defect detection and inspection