Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 33 stars

thudzj/ScalableBDL

Code for "BayesAdapter: Being Bayesian, Inexpensively and Robustly, via Bayeisan Fine-tuning"

Overview: This project provides a PyTorch implementation for Bayesian fine-tuning, allowing practitioners to learn Bayesian Neural Networks (BNNs) at low cost. Key features: - Bayesian fine-tuning for learning BNNs - Deterministic pre-training using MAP estimation - Off-the-shelf pre-trained models for deterministic pre-training - Conversion of DNN to BNN for Bayesian fine-tuning - Support for mean-field Gaussian as approximate posterior - Flexible distribution support under development Tech stack: Python Why it matters: This project offers a practical solution for learning BNNs, making Bayesian inference accessible without requiring high computational resources, thus democratizing Bayesian neural network learning.

Shell 33 stars

va-stepanov/vosk-model-ru-adaptation

Overview: This project aims to adapt and expand the Vosk model for Russian speech recognition, improving its accuracy and performance. Key features: - Adaptation of the Vosk model for Russian language - Updating the model's lexicon and grammar - Using a Docker container for easy management and monitoring Tech stack: Shell Why it matters: The project enhances the Vosk model for Russian speech recognition, improving its accuracy and performance, making it more suitable for Russian-speaking applications. Repository: va-stepanov/vosk-model-ru-adaptation Primary language: Shell GitHub stars: 33

Python 33 stars

visinf/lnfmm

Latent Normalizing Flows for Many-to-Many Cross Domain Mappings (ICLR 2020)

Overview: This repository contains the PyTorch implementation of the paper "Latent Normalizing Flows for Many-to-Many Cross Domain Mappings" presented at ICLR 2020. The code is designed for training and evaluating models on tasks such as image captioning and text-to-image synthesis. Key features: - Latent Normalizing Flows for Many-to-Many Cross Domain Mappings - PyTorch implementation - Supports Python 3.5, Pytorch 1.0.0, and CUDA 10.0 - Training script `train.py` for both image captioning and text-to-image synthesis tasks - Dataset used: COCO 2014 - CUDA 10.0 and GPU devices required for training - Checkpoints available for text-to-image synthesis and image captioning Tech stack: Python, Pytorch, CUDA 10.0 Why it matters: This repository provides a practical solution for researchers and developers interested in natural language processing and computer vision tasks, offering a robust framework for training and evaluating models on cross-domain mappings.

Python 33 stars

willxxy/ECG-Byte

[MLHC 2025] ECG-Byte: A Tokenizer for End-to-End Generative Electrocardiogram Language Modeling

Overview: This repository contains the official implementation of ECG-Byte, a tokenizer designed for end-to-end generative electrocardiogram language modeling. The project was accepted at the Machine Learning for Healthcare Conference 2025 and is part of the ECG-Bench project. Key features: - Tokenizer for ECG data - End-to-end training pipeline - Pretraining and finetuning stages - Attention visualization analysis Tech stack: Python Why it matters: ECG-Byte is a significant contribution to the field of electrocardiogram language modeling, offering a new approach to generating ECGs from text inputs. Repository: willxxy/ECG-Byte GitHub stars: 33 GitHub description: [MLHC 2025] ECG-Byte: A Tokenizer for End-to-End Generative Electrocardiogram Language Modeling

Python 33 stars

yandex-research/adaptive-diffusion

[CVPR'2024] Adaptive Teacher-Student Collaboration for Text-Conditional Diffusion Models

### Overview: In this work, we investigate the relative quality of samples produced by the teacher text-to-image diffusion model and its distilled student version. As our main empirical finding, we discover that a noticeable portion of student samples exhibit superior fidelity compared to the teacher ones, despite the "approximate" nature of the student. Based on this finding, we propose an adaptive collaboration between student and teacher diffusion models for effective text-to-image synthesis. Specifically, the distilled model produces the initial sample, and then an oracle decides whether it needs further improvements with a slow teacher model. ### Key features: - Adaptive refinement strategy using the rollback value (σ) to control the amount of Gaussian noise applied to the student sample. - Adaptive selection of samples using the cut-off threshold (τ) based on the quality estimator (IR). - Two adaptive strategies: refinement and regeneration. ### Tech stack: Python ### Why it matters: Our approach operates with the following entities: 1. Models: student, teacher, and quality estimator. 2. Hyperparameters: rollback value (σ) and cut-off threshold (τ). The algorithm below presents the proposed pipeline, consisting of two adaptive strategies: refinement and regeneration. - The rollback value (σ) is used in the adaptive refinement strategy and controls the amount of Gaussian noise applied to the student sample. It can be varied between 0 and 1. Higher σ leads to more pronounced changes but requires more teacher steps to produce plausible samples. - The cut-off threshold (τ) is necessary to perform adaptive selection. Specifically, we define it as a k-th percentile of the

JavaScript 32 stars

12we21/cute-fat-fish-pet

蓝毛小女仆 (Blue-Haired Little Maid) · 住在 Windows 桌面上的 Q 版蓝发小女仆桌面宠物：106 个手绘透明动画、本地 Ollama 或在线模型陪聊、可选看屏幕吐槽、离线语音识别（安装器与便携版）。免费开源 MIT。同时是 DeepSeek Harness 桌面宠物插件（dsh-plugin）。By Mikolu.

Overview: Blue-Haired Little Maid is a free, open-source desktop pet for Windows, featuring 106 hand-drawn animations and local or online chat capabilities. It can open apps, files, folders, and web pages, and supports offline voice recognition. Key features: - 106 hand-drawn animations - Local or online chat - Screen watching and screen recording - Data privacy and build from source Tech stack: JavaScript Why it matters: Blue-Haired Little Maid is a unique and fun desktop pet that offers a delightful user experience, making it a great addition to any Windows environment.

JavaScript 32 stars

1599570912/IELTS-Speaking-AI

🎤 AI-powered IELTS speaking practice: real-time speech recognition, scoring, 210+ questions, local history

Overview: This project is an open-source education tool designed for non-native English learners to practice IELTS Speaking (Parts 1–3) in the browser. It includes real-time speech recognition, AI scoring, and a comprehensive question bank with 210+ questions covering all parts of the exam. The application is built using HTML/JS, Web Speech API, and local storage for privacy-first data handling. It also includes an experimental language-learning prototype and scripts for OpenAI roleplay. Key features: - 🎯 Intelligent Speech Recognition: High-quality audio recording with support for various formats. - 🗣️ Real-time Speech Recognition: Web Speech API for real-time transcription. - 🤖 AI Scoring: Based on the IELTS official rubric, providing a comprehensive 4D scoring system. - 📚 Professional Question Bank: 210+ practice questions covering all parts of the exam. - 📊 Data Analysis: Detailed voice statistics and improvement suggestions. - 📝 Historical Record: Complete practice history and progress tracking. Tech stack: HTML/JS, Web Speech API, IndexedDB, Web Storage. Why it matters: This tool offers a practical and effective way for learners to improve their IELTS Speaking skills, making it a valuable resource for those preparing for the exam. Repository: 1599570912/IELTS-Speaking-AI GitHub stars: 32 Primary language: JavaScript GitHub description: 🎤 AI-powered IELTS speaking practice: real-time speech recognition

TypeScript 32 stars

AASHISHAG/DeepSpeech-API

The code enables users to use Mozilla's Deep Speech model over the Web Browser.

### Overview: This project, DeepSpeech-API, enables users to access Mozilla's DeepSpeech model directly over the web browser, facilitating speech-to-text functionality without the need for installation. ### Key features: - **Web Browser Access**: The project allows users to use Mozilla's DeepSpeech model over the web browser, making it accessible on any platform. - **Python Bindings**: Provides Python bindings for the DeepSpeech model, enabling easy integration into web applications. - **Pre-trained Model**: Offers pre-trained English model for speech-to-text functionality, reducing the need for manual model training. ### Tech stack: TypeScript ### Why it matters: This project simplifies the use of Mozilla's DeepSpeech model, making it accessible and easy to integrate into web applications, enhancing the capabilities of speech-to-text functionalities without the need for complex installation processes.

Unknown 32 stars

Claytro24161ll/telegram-channel-rpocket-forwarder-ai

Automated Telegram channel forwarder and AI content rewriter with zero latency for Windows.

Overview: This project automates the forwarding of Telegram channels with zero latency, utilizing AI for content rewrites, making it ideal for developers looking to streamline their communication and content management. Key features: - Automated Telegram channel forwarder - AI content rewriter with zero latency Tech stack: Python Why it matters: This project offers a solution that enhances communication efficiency and content quality, making it a valuable tool for developers working with Telegram channels.

Python 32 stars

ComfyUI-Kelin/ComfyUI-LLMs-Toolkit

ComfyUI custom nodes for DeepSeek, Qwen, GPT, and other OpenAI-compatible LLM APIs, with tools for chat, translation, vision, and JSON workflows.

Overview: ComfyUI-LLMs-Toolkit is a ComfyUI custom node pack designed for connecting various OpenAI-compatible LLM APIs, including DeepSeek, Qwen, GPT, and others, directly within the workflow. It includes built-in tools for chat, translation, vision, and JSON workflows, ensuring seamless integration without the need for local GPU models. Key features: - Built-in Provider Manager for easy API management. - 12 pre-configured providers: Qwen, DeepSeek, GLM, Doubao, Spark, Moonshot, Baichuan, MiniMax, StepFun, SenseChat, iFlow, ModelScope. - Smart Model Dropdown for automatic model selection. - Vision support for image processing with the Image Preprocessor node. - Error handling to prevent workflow crashes. - Multi-turn memory for chat-style interactions. Tech stack: Python Why it matters: This toolkit simplifies the integration of LLM-powered workflows into ComfyUI, making it easier to automate text and vision tasks without the complexity of local GPU model management.

Jupyter Notebook 32 stars

GalaxyXieyu/RAG-LEARN

A structured learning repo for retrieval-augmented generation, from foundations to production patterns.

### Overview: This repository is designed as a comprehensive learning map for retrieval-augmented generation (RAG), covering data cleaning, retrieval, answer generation, graph RAG, vector database architecture, and multimodal directions. It aims to provide a systematic approach from foundational concepts to enterprise-level knowledge management. ### Key Features: - Data Cleaning and Preprocessing - Retrieval Techniques - Answer Generation with LLMs - Graph RAG (Graph-based Retrieval-Augmented Generation) - Enterprise Vector Database Architecture - Multimodal Embedding Techniques ### Tech Stack: Jupyter Notebook ### Why It Matters: This repository serves as a structured learning path for those interested in RAG, providing a clear progression from basic concepts to enterprise-level knowledge management. It includes practical examples and real-world solutions to common challenges in RAG implementation. ### Repository: GalaxyXieyu/RAG-LEARN - Primary Language: Jupyter Notebook - GitHub Stars: 32 - GitHub Description: A structured learning repo for retrieval-augmented generation, from practical foundations to enterprise-scale retrieval systems.

HTML 32 stars

GeminiAtHell609142-alt/MAX2-AI-Content-Automation-resocket

Open-source human-in-the-loop AI content automation with n8n, Notion, Telegram, AI generation, and LinkedIn publishing.

Overview: This project automates human-in-the-loop AI content creation, integrating n8n, Notion, Telegram, AI generation, and LinkedIn publishing. Key features: - AI-generated content creation - Integration with n8n for process automation - Notion-based content management - Telegram notifications for updates - AI-powered content generation - LinkedIn publishing Tech stack: HTML Why it matters: This project streamlines content creation, improving efficiency and reducing manual effort while maintaining quality through AI-driven automation.

Python 32 stars

IlyasMoutawwakil/py-txi

A Python wrapper around HuggingFace's TGI (text-generation-inference) and TEI (text-embedding-inference) servers.

### Overview: Py-TXI is a Python wrapper around HuggingFace's TGI (text-generation-inference) and TEI (text-embedding-inference) servers, enabling seamless integration of these powerful models into Python applications. ### Key Features: - **Easy to Use**: Designed to be used similarly to the Transformers API. - **Automatic Cleanup**: Stops the Docker container automatically when your code finishes or fails. - **Batched Inference**: Supports sending a batch of inputs for inference. - **Automatic Port Allocation**: Automatically allocates a free port for the Inference server. - **Configurable**: Allows configuring the Inference servers using a simple configuration object. - **Verbose Logs**: Streams logs of the underlying Docker container to the main process for easy debugging. ### Tech Stack: - Python ### Why It Matters: - **Ease of Use**: Simplifies the integration of TGI and TEI models into Python applications. - **Automatic Cleanup**: Ensures that the Docker containers are stopped automatically, freeing up resources. - **Efficiency**: Supports batched inference and automatic port allocation, enhancing performance and efficiency. - **Configuration Flexibility**: Allows for easy configuration of the Inference servers. - **Debugging**: Provides verbose logs for easy debugging and troubleshooting. ### GitHub Stars: 32 ### Primary Language: Python ### GitHub Description: A Python wrapper around Text-Generation-Inference and Text-Embedding-Inference that enables creating and running TGI/TEI instances through the awesome `docker-py` in a similar style to Transformers API.

Rust 32 stars

IronCoreLabs/ironcore-alloy

A unified SDK for IronCore encryption operations including vector encryption for AI data

### Overview: The IronCore Labs Alloy SDK provides a unified set of tools for application layer encryption, including vector encryption for AI data. It simplifies the process of protecting sensitive information across various data sources such as structured documents, relational databases, key-value stores, and vector databases. ### Key Features: - **Unified Encryption**: Simplifies the process of encrypting data across different types of storage and application layers. - **Vector Encryption**: Specifically designed for AI data, ensuring secure handling of embeddings and vectors. - **Standard, Deterministic, and Vector Encryption**: Supports various encryption algorithms for different use cases. ### Tech Stack: Rust ### Why It Matters: The Alloy SDK offers a comprehensive solution for encrypting data, making it easier to secure sensitive information in various applications and storage systems. By unifying encryption tools, it streamlines the process and ensures that sensitive data is protected effectively without requiring extensive configuration or additional libraries. ### Repository: Repository: IronCoreLabs/ironcore-alloy GitHub Stars: 32 GitHub Description: A unified SDK for IronCore encryption operations including vector encryption for AI data GitHub Mentioned in: Awesome Security Solutions for AI Systems GitHub Version: 0.1.0 ### Getting Started: - Follow the links provided to get the latest version for the appropriate language. - Explore examples on the main documentation site [here](https://ironcorelabs.com/docs/). - For Cloaked AI integration examples, refer to the specific vector database integration examples [here](https://ironcorelabs.com/docs/cloaked-ai/integr

Python 32 stars

JustVugg/distillery

Turn documents into high-quality instruction datasets with grounding, quality filtering, deduplication, and provenance.

### Overview: Distillery is a Python library and CLI designed to turn documents into high-quality instruction datasets. It ensures every example is grounded, graded by an LLM judge, deduplicated by instruction embedding, and has provenance, making it suitable for fine-tuning models. ### Key features: - Grounded examples - Graded by an LLM judge - Duplicated examples by instruction embedding - Provenance tracking - Exported in multiple formats ### Tech stack: Python ### Why it matters: Most teams fine-tune on garbage datasets because producing a clean one is manual work. Distillery automates the loop, resulting in a dataset that is actually usable, with a `manifest.json` that records exactly what went in, what was rejected, and why.

Shell 32 stars

KaiFelixBennett/hermes-claude-code-local

Run Hermes Agent + Claude Code locally on llama.cpp — zero API costs. A 4h / 7M-token session that would have cost $94 on Claude Opus 4.7

Overview: This project provides a local setup for the Hermes Agent and Claude Code, allowing for agentic coding without the need for an API. The setup includes llama.cpp as the inference engine and optionally a LiteLLM proxy for Claude Code, resulting in zero API costs. Key features: - Local agentic coding - Hermes Agent running with llama.cpp - Claude Code bridge via LiteLLM - Telegram integration - Zero API costs - MTP / speculative decoding - Self-healing scripts - Model tuning notes Tech stack: Shell Why it matters: This setup offers a cost-effective alternative to the Claude Opus API, providing a 4-hour session equivalent to $0, making it ideal for developers seeking to avoid API costs. Repository: KaiFelixBennett/hermes-claude-code-local GitHub stars: 32

Jupyter Notebook 32 stars

Kirili4ik/ruDialoGpt3-finetune-colab

Ready-for-use-colab tutorial for finetuning ruDialoGpt3 model on а telegram chat using HuggingFace and PyTorch

Overview: This is a ready-for-use Colab tutorial designed for fine-tuning the ruDialoGPT3 model on a telegram chat using HuggingFace and PyTorch. Key features: - Fine-tuning ruDialoGPT3 model on telegram chat - Colab tutorial ready for use - Model trained on Russian forums by Grossmend - Simple pipeline fine-tuned on exported telegram chat (~30MB json, 3 hours of fine-tuning) Tech stack: Jupyter Notebook Why it matters: This tutorial provides a practical guide for developers looking to fine-tune a model on real-world data, making it easier to adapt and improve language models for various applications.

C++ 32 stars

MChatzakis/DARTH

[SIGMOD 2026] DARTH: Declarative Recall Through Early Termination for Approximate Nearest Neighbor Search.

Overview: DARTH is a method that uses target declarative recall to improve the performance and recall of approximate nearest neighbor search (ANNS) algorithms, offering a more user-friendly approach to tuning parameters for specific recall targets. Key features: - DARTH provides declarative recall targets for ANNS algorithms. - It integrates an adaptive early termination strategy into the search algorithm. - It achieves significant speedups and improved recall compared to traditional ANNS methods. Tech stack: C++ and FAISS library. Why it matters: DARTH simplifies the process of tuning ANNS parameters, offering a more user-friendly approach to achieving specific recall targets, thereby improving the performance and recall of search operations.

Python 32 stars

MarcosNahuel/antigravity-plugin-cc

Local NotebookLM for Claude Code via Google Antigravity (agy / Gemini 3.x): /agy:notebook turns a folder of documents into per-doc summaries + a relevance index + a cited synthesis + Q&A. Plus audio/video transcription, deep web research with citations & branded HTML reports. 13 commands, no Node runtime.

### Overview: Antigravity Plugin for Claude Code is a powerful local NotebookLM that leverages Google Antigravity CLI to transform documents into summaries, indexes, and more. It includes features like audio/video transcription, deep web research with citations, and branded HTML reports, all without requiring Node.js or any MCP gymnastics. The plugin is designed to be used with Claude Code, replacing the deprecated Gemini CLI, and is compatible with Enterprise users on Code Assist Standard/Enterprise licenses. ### Key features: - **Local NotebookLM**: Converts folders of documents into per-document summaries, relevance index, cited synthesis, and timeline. - **Audio/Video Transcription**: Transcribes audio and video content. - **Deep Web Research**: Runs multi-agent deep research with evidence matrix and plan approval. - **Branded HTML Reports**: Generates branded HTML reports. - **Git-Diff Code Review**: Performs git-diff code review. - **Doc-to-Markdown**: Converts documents to markdown. - **Browser Recording**: Records browser actions. - **Multi-Agent Deep Research**: Adaptive rounds with Claude orchestrating rounds (depth L<=2, H<=4) and a red-team pass. - **No Node.js Runtime**: No need for Node.js runtime. - **22 Slash Commands**: 22 commands for easy usage. ### Tech stack: The plugin uses Python and Google Antigravity CLI to achieve its functionality. ### Why it matters: This plugin is essential for anyone looking to streamline their Claude Code workflow, especially for enterprise users who need to replace the deprecated Gemini CLI. It provides

JavaScript 32 stars

Mdx2025/BrainX-The-First-Brain-for-OpenClaw

🧠 BrainX V6 — The First Brain for OpenClaw. Persistent AI agent memory with PostgreSQL, pgvector, OpenAI embeddings, semantic search, cross-agent learning, and an OpenClaw runtime plugin.

Overview: BrainX V6 is a persistent AI agent memory system designed for OpenClaw, providing long-term memory, semantic search, cross-agent learning, and an OpenClaw runtime plugin. It uses PostgreSQL, pgvector, and OpenAI embeddings for persistent vector storage and auto-learning capabilities. Key features: - Persistent vector storage - Auto-learning - Persistent memory - Cross-agent learning - Runtime plugin Tech stack: JavaScript Why it matters: BrainX V6 offers a robust, scalable, and safe AI system for OpenClaw agents, ensuring long-term memory and knowledge sharing across sessions, reducing redundancy, and improving decision-making through semantic search and cross-agent learning. Repository: Mdx2025/BrainX-The-First-Brain-for-OpenClaw GitHub stars: 32 GitHub description: 🧠 BrainX V6 — The First Brain for OpenClaw. Persistent AI agent memory with PostgreSQL, pgvector, OpenAI embeddings, semantic search, cross-agent learning, and an OpenClaw runtime plugin.