Physically conditioned correction and ordered uncertainty for multiday particulate forecasting over East Asia — PM2.5/PM10 exceedance, +6h to +120h

Solar filament segmentation using U-Net architectures in TensorFlow/Keras, benchmarking custom CNN baseline vs. EfficientNetB0 transfer learning and fine-tuning.

Mistral Timestral: Tiny 209M AI Predicts Stocks Locally on CPU! Zero-shot probabilistic forecasting foundation model.

Liquid Neural Network language model with LTC and CfC dynamics

Handwritten digit recognition with a CNN trained on MNIST (99.32% accuracy). Upload or snap a photo and get the predicted digit with confidence and full preprocessing visualization. Built with TensorFlow, OpenCV, and Streamlit.

AI-Powered Bird Biodiversity & Environmental Monitoring Platform (PyTorch + FastAPI + Supabase + React)

Hands-on Jupyter notebooks teaching MRI reconstruction: from k-space and the Fourier transform to compressed sensing and deep learning. Runs in Google Colab with simulated data.

PROJECT LOOMY: Official reference implementation of Loomy (luminamoon) - Low-precision BF16/FP16 deep learning optimizer (LuminaV CLM & TD-CLM), high-throughput multimodal data processing, and formal Z3 SMT verification in Rust & Python.

🤖 AI model that measures water purity from microscopy images via deep-learning Trophic Diatom Index (TDI) estimation. Crafted in Spain 🇪🇸.

2-bit LLMs for llama.cpp: 85% smaller, beats IQ2_XXS at the same size, no retraining. One-click Windows build.

Implementation of a deepfake forensics pipeline combining Fourier spectral analysis, dense optical flow, WavLM audio representations, image-artifact profiling, and vision-language verification with calibrated ensemble scoring.

GateFormer: forecast-space decomposition for intra-hour solar irradiance forecasting from sky images. Energy and AI (2026). Paper, results and citation. Code coming soon.

Self-hosted AI experiment workbench for SSH GPU servers: presets, code snapshots, queued runs, project-defined checkpoint resume, and offline charts and comparisons. Desktop apps for macOS, Windows and Linux, plus CLI and an iPhone Web App.

Awesome AI Research Tools (2026): SOTA catalog of literature discovery, citation synthesis, LaTeX, and reproducible MLOps

AI for the Flash/AIR fighting game Bleach vs Naruto (BVN): injected ActionScript state bridge, behaviour cloning + DAgger + conservative offline RL, verified against the real game

Automated dermatological lesion classification on HAM10000 in PyTorch. Paper reproduction (BMC Medical Imaging 2024) and EfficientNet-B0 transfer learning benchmark.

A transformer that is a working computer. It runs CHIP-8 games like Tetris and Pong by predicting every write to its own memory, one token at a time.

A desktop SSH console for Google Colab runtimes: a real terminal, the runtime's files over SFTP, and the compute units it is spending. Unofficial client.

On-device tennis ball tracking: a TrackNet-style CNN in PyTorch, evaluated leakage-free and deployed to Apple Neural Engine via Core ML (52 FPS).

Official PyTorch implementation of FLEX-MRI: Stochastic Finite-Transition Learning and Exploration for Accelerated MRI Reconstruction

Real-time AI-assisted freehand knee ultrasound for standard-plane classification and multi-structure segmentation

KV cache compression for LLM inference: 37× smaller KV cache at full-cache accuracy. Per-token precision on transform coding, question-time refinement, and attention over the compressed cache. PyTorch + Hugging Face.

Multi-agent match intelligence built on synthetic football event data (no real Premier League data used). Built for the Inside the Game: Developer Hackathon.

An interactive map of the whole AI field across its history: a timeline, a family tree and today's products in one view.

End-to-end robotic-arm imitation learning with LeRobot × SO-ARM101: teleoperation data collection, ACT training, real-world deployment, and a root-cause analysis of behavior-cloning overfitting.

How self-supervised learning works: a masked autoencoder finds brain lesions with zero labels and beats training from scratch with few labels

IEFG-CDAN: Information-Entropy-aware Fine-Grained Conditional Domain Adversarial Network for cross-working-condition bearing fault diagnosis (CWRU)

Vision-based object sorting (RealWaste) for the Computer Vision midterm-to-final 4WD robot project

Collaborative research at the intersection of breast histopathology, data science, and deep learning.

Run very large MoE models (100B–1T+) on the hardware you already own, by streaming experts across VRAM/RAM/SSD. Source-available, free for personal use.