Agentic AI system for brain MRI analysis: segmentation, uncertainty estimation, vision-language report generation, and automatic human-review flagging.

A synthesizable FlashAttention engine in SystemVerilog: verified bit-exact against a golden model, compiled to WebAssembly, running live in your browser.

Hands-on implementations of ML, DL, NLP, and CV models and architectures from scratch and with modern frameworks.

Adversarial ML from scratch in NumPy: gradient-checked models, FGSM/PGD (targeted) attacks, adversarial training (35% -> 77% robust accuracy) and a manifold detector vs feature squeezing, with honest results.

Deep learning pipeline for lung cancer detection (YOLOv8, mAP@0.5 = 0.75) and NSCLC vs SCLC classification (ResNet18) on CT scans from the public Lung-PET-CT-Dx dataset. Final degree project, Biomedical Engineering (Universidad de Málaga).

AI study roadmap 2026: learn AI engineering by building — one track, stage by stage, zero spend on LLM APIs

Top Distributed Training Optimization (Opensource) Star if you like it!

A config-as-graph framework for deep learning: any pipeline, in any modality or domain, assembled from swappable components.

A programming language for AI and data work with plain-English syntax, a tensor engine with autograd, a Python bridge, and a package manager.

Andrew Ng Deep Learning, an educational AI resource designed to break down foundational neural network concepts, machine learning algorithms, and model optimization techniques. It provides a seamless user experience to help students and AI practitioners build strong deep learning fundamentals and master core data science principles.

Interactive visual guide to LLM and generative AI architectures — attention variants, MoE, state-space models, fine-tuning and diffusion, shown tensor by tensor.

Offline two-stage encoding-model selection and MEI generation for pooled single-photon neural responses.

Enter the era of Function-Aware Optimization

ANN-based estimated salary prediction system built with Python, TensorFlow, Scikit-learn, and Streamlit.

A small character-level GPT in PyTorch (LayerNorm, dropout, AdamW) with a CLI for training, sampling, and loss-curve evaluation.

Network intrusion detection with PyTorch ANNs on UNSW-NB15 (10-class) with confidence-based alerting

CASIA internship-derived traffic forecasting: Random Forest + PyTorch GRU, reproducible temporal evaluation, 12.4% lower GRU MAE, FastAPI inference.

Clinical 2D/3D biomedical image segmentation with U-Net, Soft Dice Loss, and ONNX Runtime PACS export (2016-2018).

Waveform-domain speech denoising with a 230k-parameter 1-D U-Net, plus a Gradio app that transcribes before and after with Whisper.

Benign vs malignant skin lesion classification from dermoscopic images: custom CNN vs fine-tuned ResNet18. 90.0% accuracy, 0.967 AUROC.

Deep Learning Projects that I created towards end of my Bachelors

ANN-based customer churn prediction system built with Python, TensorFlow, Scikit-learn, and Streamlit.

ML-powered Wi-Fi security research platform for anomaly detection, network behavior analysis and authorized wireless security experiments.

A small GPT that writes molecules as SMILES strings, with an interface to explore them in 2D and 3D. Early version 0.1, more coming.

Protocol outweighs method: patch sampling, evaluation scope, consensus level and threshold dominate architecture and supervision-target choice in lung nodule segmentation on LIDC-IDRI.

🌾 AI-Powered Crop Disease Detection & Agronomy Advisory Platform for Farmers. Fine-tuned MobileNetV2 on PlantVillage (38 classes), bilingual Hindi (हिंदी) & English guidance, camera leaf scanner, and organic/chemical remedies.

Guía técnica en español de las arquitecturas de IA: qué son, para qué sirven y cómo funcionan, con ejemplos en PyTorch

Deterministic, high-performance Rust deep learning engine and framework with zero silent fallback, in-process Go bindings, and strict memory budgeting

BoundCertain is a deep learning project for pixel level scene understanding in street images. It combines boundary aware feature fusion with uncertainty estimation to improve scene segmentation and help identify potentially unreliable predictions.

LSTM baselines for A-share stock crash prediction: dual crash labels (extreme-return and market-model crash weeks), weekly panels (40-1577 stocks), reproducible training/eval pipeline.