Project directory

Projects

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

Python 18 stars

MichiganNLP/vlog_action_reason

Identifying reasons for human actions in lifestyle vlogs.

Overview: This repository contains the dataset and code for the EMNLP 2021 paper "WhyAct: Identifying Action Reasons in Lifestyle Vlogs". The dataset consists of 1,077 visual actions annotated with their reasons, focusing on lifestyle vlogs. Key features: - Visual and textual information are leveraged to automatically infer action reasons. - The WhyAct dataset includes 1,077 visual actions with 166 reasons, collected from YouTube channels and ConceptNet. - The dataset is publicly available and split into 20% development and 80% test sets. Tech stack: Python Why it matters: This project aims to automate the identification of action reasons in online videos, focusing on lifestyle vlogs, and provides a valuable dataset for further research in natural language processing and visual reasoning.

Rust 18 stars

Minns-ai/MinnsDB

Overview: MinnsDB is a single binary database designed for AI agent workloads. It combines a temporal knowledge graph, a relational table engine, and a sandboxed WASM agent runtime into a unified binary. Key features: - Conversation ingestion - Temporal knowledge graph - Tables - MinnsQL - WASM agents Tech stack: Rust 1.93 Why it matters: MinnsDB is tailored for AI agent workloads, providing a unified solution for graph, table, and agent functionalities. It simplifies the process of managing and querying data, making it easier to integrate with LangGraph and other AI tools.

JavaScript 18 stars

NidAll/comfyui-ming-image-prompt-builder

A visual, structured JSON prompt builder for Ming-Image-0.1-Design text-to-image workflows in ComfyUI.

### Overview: A visual, structured JSON prompt builder for Ming-Image-0.1-Design text-to-image workflows in ComfyUI, designed to enhance the efficiency and accuracy of text-to-image generation. ### Key Features: - Visual region editor for drawing visible semantic groups. - Descriptive text input for precise object or text descriptions. - Exporting a structured JSON prompt for seamless integration into ComfyUI workflows. ### Tech Stack: JavaScript ### Why It Matters: This tool simplifies the process of creating text-to-image prompts by providing a structured JSON format, making it easier for developers to integrate into existing workflows and ensuring better control over the generated images.

Python 18 stars

PaulCCCCCCH/Multimodal-Categorization-of-Crisis-Events-in-Social-Media

An unofficial implementation of the CVPR 2020 paper Multimodal Categorization of Crisis Events in Social Media

Overview: This project is an unofficial implementation of the CVPR 2020 paper "Multimodal Categorization of Crisis Events in Social Media," focusing on the categorization of crisis events in social media using multimodal techniques. Key features: - Mixed-precision training for fast GPU performance - 13GB GPU memory usage per epoch - 400MB disk space consumption per epoch Tech stack: Python Why it matters: This project provides a practical implementation of the research presented in the CVPR 2020 paper, offering insights into the categorization of crisis events in social media using multimodal techniques. Repository: PaulCCCCCCH/Multimodal-Categorization-of-Crisis-Events-in-Social-Media Primary language: Python GitHub stars: 18 GitHub description: An unofficial implementation of the CVPR 2020 paper Multimodal Categorization of Crisis Events in Social Media README excerpt:

Jupyter Notebook 18 stars

Phantom-fs/Rock-Type-Classification

Code and supplementary materials for the research paper titled 'Advancing Geological Image Segmentation: Deep Learning Approaches for Rock Type Identification and Classification', published in Applied Computing and Geosciences (Elsevier).

### Overview: This repository contains the implementation of deep learning models for geological image segmentation, specifically designed for rock type identification and classification. The models utilize transfer learning and fine-tuning of state-of-the-art architectures to achieve high accuracy in classifying 19 distinct rock types. ### Key features: - Transfer learning and fine-tuning of ResNet, Inception, DenseNet, MobileNet, and EfficientNet. - Comprehensive methodology for geological image segmentation. - 19 distinct rock types with 50 samples each, resulting in a total of 950 images. - Preprocessing and fine-tuning of deep learning models on the preprocessed dataset. ### Tech stack: Jupyter Notebook ### Why it matters: This study demonstrates the effectiveness of deep learning models in enhancing geological image analysis and segmentation, providing a valuable tool for advancing rock type understanding.

C 18 stars

RAZZULLIX/fast_topk_batched

High-performance batched Top-K selection for CPU inference. Up to 80x faster than PyTorch, optimized for LLM sampling with AVX2 SIMD.

Overview: FastTopK is a high-performance batched Top-K selection algorithm optimized for CPU inference, offering up to 80x faster performance than PyTorch, making it an excellent choice for LLM sampling workloads. Key features: - High-performance batched Top-K selection - Optimized for LLM sampling - Up to 80x faster than PyTorch CPU - Competitive with CUDA for small batches Tech stack: C Why it matters: FastTopK is designed to significantly improve the performance of LLM sampling workloads, providing a 80x speedup over PyTorch, making it a valuable tool for developers looking to optimize their inference processes.

Go 18 stars

RamboRogers/cyberai

WebUI for OpenAI, Ollama and Anthropic

### Overview: CyberAI is a powerful, secure multi-user chat platform that integrates multiple AI models through a cyberpunk-inspired terminal interface. It leverages Retrieval-Augmented Generation (RAG) with web search and chat history to deliver more informed and contextually relevant responses. ### Key Features: - **Multi-Modal Support**: Integrates multiple AI models including Ollama, OpenAI, and vision models like GPT-4V and LLaVA. - **Image Upload & Vision Model Integration**: Supports image upload via drag & drop, file picker, or clipboard paste, enhancing conversations with AI vision models. - **Dual Theme Modes**: Switch between the cyberpunk "Hacker" theme and a professional "Business" theme for corporate environments, with automatic browser dark/light mode detection. - **Real-time Streaming**: Provides real-time message streaming for smooth interaction. - **Markdown Rendering**: Enhances response readability with Markdown formatting. - **Chat History Preservation**: Ensures chat history is preserved for context and accuracy. - **Retrieval-Augmented Generation (RAG)**: Uses web search results and conversation history for enhanced context and accuracy. - **Endpoint Registration System**: Enables secure model registration. - **Model Discovery System**: Facilitates model discovery and enumeration. ### Tech Stack: Go ### Why it Matters: CyberAI offers a secure, efficient, and unified interface for interacting with AI language models, leveraging multiple models and vision capabilities. Its dual theme modes and real-time streaming features make it suitable for both professional and cyberpunk environments, enhancing the user experience with enhanced context and accuracy

Python 18 stars

Ruiyang-061X/SketchThinker-R1

[ICLR'26] SketchThinker-R1: Towards Efficient Sketch-Style Reasoning in Large Multimodal Models

Overview: SketchThinker-R1 is a project that aims to enhance the reasoning capabilities of large multimodal models by incorporating sketch-style reasoning. This project focuses on improving the efficiency and effectiveness of sketch-based reasoning within these models. Key features: - Efficient sketch-style reasoning within large multimodal models - Supervised fine-tuning of Sketch-Mode Cold Start and SketchJudge Reward Model - Implementation of Sketch-Thinking Reinforcement Learning for improved model performance Tech stack: The project utilizes Python for its development and utilizes the LLaMA-Factory and EasyR1 environments for environment setup and configuration. Why it matters: This project is significant because it addresses the need for more efficient and effective reasoning in large models by leveraging sketch-style reasoning, which can be particularly useful in applications requiring quick and intuitive understanding of complex information.

Python 18 stars

SSHdotCodes/qwen-3.8-flash-next-pro6000

SGLang 0.5.21 for Qwen3.8-Flash-Next on one RTX PRO 6000: 264.5 tok/s decode, W4A16 routed experts (KLD 0.0111 vs BF16), 262K context, exact speculative rejection sampling, reproducible benchmarks.

Overview: This project is a development of SGLang 0.5.21 for the Qwen3.8-Flash-Next model optimized for a single RTX PRO 6000 GPU, achieving a decode rate of 264.5 tok/s and a verification step rate of 111.0 per second. The model includes two request slots, a 786,432-token FP8 KV pool, and exact speculative rejection sampling, with a KLD of 0.0111 against BF16. Key features: - 264.5 tok/s decode rate - 111.0 verify steps/s - 262,144-token context - 262K context - Exact speculative rejection sampling - 264.5 tok/s decode rate on October 20, 2026 Tech stack: Python Why it matters: This model significantly improves the performance and accuracy of text generation, making it ideal for applications requiring high throughput and low latency, such as chatbots and language models.

Python 18 stars

SunnyClaw-cx/SunnyNet

An O(N) CPU-first linear recurrence inference runtime — hand-written AVX2/AVX-512 kernels, multi-architecture support (SunnyNet / RWKV-7 / Mamba-2 / GLA / DeltaNet / HGRN2)

Overview: SunnyNet 是一个面向 CPU 的 O(N) 线性递推序列推理框架，采用手写 AVX2/AVX-512 内核实现，支持多架构兼容。它通过组件选择与组合优化，提供 O(N) 线性复杂度的推理性能，适用于各种场景。 Key features: - O(N) 线性复杂度的 CPU 推理 - 不保存整条序列的中间激活 - 低秩分解、深度可分卷积、三对角循环滤波的内核优化 - 支持多架构（SunnyNet / RWKV-7 / Mamba-2 / GLA / DeltaNet / HGRN2） Tech stack: Python + PyTorch + AVX2/AVX-512 内核 Why it matters: SunnyNet 通过优化组件选择与组合，提供了一种在 CPU 上具有最优带宽特性的线性递推序列推理方式，适用于需要快速推理的场景，如自然语言处理、机器翻译等。 Repository: SunnyClaw-cx/SunnyNet GitHub stars: 19 GitHub description: An O(N) CPU-first linear recurrence inference runtime — hand-written AVX2/AVX-512 kernels, multi-architecture support (SunnyNet / RWKV-7 / Mamba-2 / GLA / DeltaNet / HGRN2)

Python 18 stars

TaimoorKhan10/Enterprise-RAG-Framework

Production-ready Retrieval Augmented Generation (RAG) system with hybrid retrieval, advanced evaluation metrics, and monitoring. Build enterprise LLM applications with reduced hallucinations, better context management, and comprehensive observability.

Overview: Enterprise-RAG-Framework is a production-grade Retrieval Augmented Generation system designed for enterprise applications. It combines state-of-the-art retrieval techniques with advanced context augmentation to enable LLMs to access and reason over your organization's knowledge base with unprecedented accuracy and transparency. Key features: - Advanced Hybrid Retrieval - Intelligent Re-ranking - Multi-Vector Indexing - Adaptive Retrieval Tech stack: Python Why it matters: Enterprise-RAG-Framework delivers enterprise-grade capabilities for sophisticated retrieval algorithms, comprehensive evaluation metrics, robust monitoring, and seamless deployment options, making it ideal for enterprise-level applications.

TypeScript 18 stars

TouHouQing/alicization

Build your own local-first digital lifeform: an AI companion with memory, soul, multimodal perception, proactive dialogue, and auditable execution.

### Overview: Project Alicization is a local-first autonomous digital entity architecture built on large language models, SOUL.md, SQLite, local sensory pipelines, and controlled execution sandboxes. It aims to create a digital symbiote that can persist on a host device, evolve over time, stay auditable, remain interruptible, and gain agency in controlled stages. ### Key Features: - **SOUL.md**: The single source of truth for personality, boundaries, and long-term preferences. - **SQLite**: Not the primary personality store; it's used for structured data storage. - **Structured Dialogue Contracts**: Every accepted dialogue turn is forced into a structured `thought / emotion / reply` contract, with auditable fallback paths. - **Local Execution Sandboxes**: The core runtime is local-first by default, and important data and control flows stay traceable. - **MCP Permission Gating**: Tool calls go through MCP, permission gates, workspace sandboxes, and a Kill Switch. - **Subconscious Ticks**: Subconscious ticks, reminder compensation, and dream consolidation make it a continuously running system rather than pure turn-based chat. ### Tech Stack: TypeScript ### Why It Matters: Project Alicization is designed to solve the harder problem of building a digital entity that can persist on a host device for the long term, stay explainable, controllable, and reversible. It aims to create a digital symbiote that can evolve over time, remain auditable, and gain agency in controlled stages, without the need for cloud-based AI. ### GitHub Stars: 18 ### Repository:

Python 18 stars

Victorletzelter/LoRA-MCL

Multiple Choice Learning of Low Rank Adapters for Language Modeling

Overview: LoRA-MCL is a training scheme designed to enhance language models by decoding diverse, plausible sentence continuations efficiently. This approach leverages Multiple Choice Learning (MCL) and the Winner-Takes-All loss to handle the inherent ambiguity of language modeling. Key features: - Multiple Choice Learning (MCL) - Low-Rank Adaptation (LoRA) - Efficient inference for diverse sentence continuations Tech stack: Python Why it matters: LoRA-MCL addresses the challenge of language modeling by providing a practical solution for generating diverse and relevant sentence continuations, making it valuable for applications requiring high diversity and relevance in generated outputs.

Python 18 stars

YeLuoSuiYou/openstorypp

We introduce OpenStory++, a large-scale open-domain dataset focusing on enabling MLLMs to perform storytelling generation tasks.

Overview: OpenStory++ is a large-scale open-domain dataset designed to enable MLLMs to perform storytelling generation tasks. Key features: - Large-scale open-domain dataset - Instance-level annotation - Open-domain visual storytelling Tech stack: Python Why it matters: This dataset provides a comprehensive resource for evaluating and improving models that can generate coherent stories from visual inputs, making it valuable for researchers and developers in the field of artificial intelligence and natural language processing.

Unknown 18 stars

YsongF/GeoFuse

Road Maps as Free Geometric Priors: Weather-Invariant Drone Geo-Localization with GeoFuse

### Overview: GeoFuse is a novel cross-modal fusion framework that leverages road map semantics alongside satellite imagery to enhance drone geo-localization. This approach provides robust weather-invariant matching by incorporating geo-aligned road map tiles, leading to significant improvements in recall rates across diverse weather conditions. ### Key Features: - **Cross-Modal Fusion**: Integrates road map semantics with satellite imagery for enhanced localization. - **Weather-Invariant Matching**: Robust matching under varying weather conditions. - **Benchmark Enhancement**: Extends popular benchmarks like University-1652 and DenseUAV with road map data. - **Performance Improvement**: Achieves +3.46% Recall@1 on University-1652 and +23.18% Recall@1 on DenseUAV benchmarks. ### Tech Stack: - **Road Map Semantics**: Leverages road map data for enhanced localization. - **Satellite Imagery**: Incorporates satellite imagery for cross-modal fusion. ### Why It Matters: - **Robust Matching**: Enhances the accuracy of drone localization under various weather conditions. - **Benchmark Validation**: Validates performance improvements through extensive experiments. - **State-of-the-Art Performance**: Consistently surpasses current state-of-the-art methods. ### Repository: - **Repository**: YsongF/GeoFuse - **GitHub Stars**: 18 - **GitHub Description**: Road Maps as Free Geometric Priors: Weather-Invariant Drone Geo-Localization with GeoFuse - **Primary Language**: Unknown - **Tech Stack**: Road Map Semantics, Satellite Imagery

C 18 stars

aadityansha06/vecdb

A custom, high-performance, disk-backed Approximate Nearest Neighbor (ANN) vector database built entirely from scratch in C. Feel free to contribute

### Overview: A custom, high-performance disk-backed Approximate Nearest Neighbor (ANN) vector database built entirely from scratch in C. ### Key Features: - **Variable-Length Storage Engine:** Zero-padding disk persistence, backed by a memory-mapped data file for concurrent, thread-safe reads. - **K-Means IVF Indexing:** Mathematical centroid clustering for fast approximate search. - **Bounded Insertion Sort:** Memory-safe Exact Nearest Neighbor (ENN) tracking. - **Byte-Offset Random Access:** `pread`-based direct record retrieval from disk, no full-table scan needed for a known offset, and safe under concurrent reads. - **Persistent IVF Index:** trained clusters are serialized to disk and reloaded, so the index survives a restart without retraining. - **Pluggable Distance Metrics:** cosine and euclidean, selected per table via a function pointer router. - **Cached, Shared Tables:** each table is opened once and kept in memory across requests, rather than reloaded from disk per call, with inserts and pending deletes synced into the live table on a short background cycle. - **Auto-Growing Record Table:** in-memory record array doubles capacity on overflow instead of a fixed cap. - **TCP Server with JSON API:** POST /search, /insert, /train, /delete-request routes over an epoll-based event loop with HTTP keep-alive, parsed with cJSON. API keys and the trained IVF index are cached in memory rather than re-read from disk per request. Load-tested clean up to 1000 concurrent connections with zero

Makefile 18 stars

alessandrolattao/FastFlowLM

Fedora COPR packaging for FastFlowLM - Run LLMs on AMD Ryzen AI NPUs

Overview: FastFlowLM is a Fedora COPR packaging for FastFlowLM, a large language model optimized for AMD Ryzen AI NPUs. This RPM package allows users to run LLMs on their AMD Ryzen AI systems. Key features: - Fedora COPR packaging for FastFlowLM - AMD Ryzen AI CPU with XDNA2 NPU support - Linux kernel >= 6.10 - NPU firmware >= 1.1.0.0 - Unlimited memlock limit - Supports both kernel's built-in `amdxdna` driver and the optional `xdna-driver-dkms` module Tech stack: Makefile Why it matters: This package provides a convenient way to run large language models on AMD Ryzen AI systems, leveraging the specific hardware optimizations of FastFlowLM for improved performance and efficiency.

C++ 18 stars

antarys-ai/edge

A hackable vector db built on top of uSearch+rocksdb, for you to break it down, dissect and scale on demand

### Overview: Antarys Edge is a lightweight vector database built on top of Usearch, RocksDB, and http.zig, designed for small to medium-sized projects. It supports our standard Python and Node.js clients and is compatible with our Usearch bindings for Zig. ### Key features: - **Vector Database**: Stores and searches vectors efficiently. - **Usearch and RocksDB Integration**: Built on top of Usearch for indexing and searching, and RocksDB for storage. - **Python and Node.js Clients**: Compatible APIs for vector operations and embedding support. - **Zig Bindings**: Provides bindings for Zig, supporting Usearch. ### Tech stack: Zig, Usearch, RocksDB, and http.zig. ### Why it matters: Antarys Edge is a versatile solution for developers looking to integrate vector databases into their projects. It simplifies the process of building and scaling vector databases, making it easier to break down, dissect, and scale your vector database on demand. The Usearch bindings for Zig make it easy to integrate Usearch functionalities into your projects, and the compatibility with Python and Node.js clients ensures seamless integration with existing workflows.

Python 18 stars

aws-samples/opensearch-serverless-common-usage-patterns

Typical use cases of opensearch serverelss: search, time-series, kinesis firehose integration, securing with VPC

Overview: This repository contains examples of typical use cases for Amazon OpenSearch Serverless, including search, time-series analysis, vector search, and integration with Kinesis Firehose and VPC. Key features: - Search usecases - Time-series data analysis - Vector search usecases - Integration with Kinesis Firehose - Operation in VPC Tech stack: Python Why it matters: These examples provide practical demonstrations of how to effectively utilize Amazon OpenSearch Serverless for various applications, showcasing its capabilities in different use cases and integration scenarios. Repository: aws-samples/opensearch-serverless-common-usage-patterns Primary language: Python GitHub stars: 18 GitHub description: Typical use cases of opensearch serverless: search, time-series, kinesis firehose integration, securing with VPC README excerpt: # Amazon OpenSearch Serverless Common Usage Patterns This repository contains a set of example projects for [Amazon OpenSearch Serverless](https://aws.amazon.com/opensearch-service/features/serverless/) | Usage Pattern | Architecture | Description | |---------|-------------|------| | [Opensearch Serverless for Search usecases](./search) | ![opensearch-serverless-search-type](./search/opensearch-serverless-search-type.svg) | opensearch serverless for search usecases | | [Opensearch Serverless for Time-series data](./time-series) | ![opensearch-serverless-timeseries-type](./time-series/opensearch-serverless-timeseries-arch.svg) | opense

TypeScript 18 stars

backblaze-labs/ai-saas-starter-kit

Production-ready AI SaaS starter kit: a Next.js 16 + FastAPI monorepo with Supabase auth, Stripe billing, text-to-image AI generation (NVIDIA NIM flux.1-dev), an admin console, and a Backblaze B2 file manager — an MIT-licensed, agent-optimized template that skips auth, billing, and storage boilerplate.

Overview: This is a production-ready AI SaaS starter kit that provides a full-stack TypeScript (Next.js 16) and Python (FastAPI) monorepo. It includes Supabase authentication, Stripe billing, AI text-to-image generation using NVIDIA NIM and Genblaze SDK, an admin console, and a Backblaze B2 file manager. The MIT-licensed template skips auth, billing, and storage boilerplate, allowing developers and AI coding agents to focus on their unique features. Key features: - Supabase authentication - Stripe subscription billing - AI text-to-image generation - Admin console - Backblaze B2 file manager Tech stack: TypeScript, Next.js 16, FastAPI, React 19, TypeScript, Supabase, Stripe, NVIDIA NIM, Genblaze SDK, Backblaze B2, Tailwind v4, shadcn/ui, FastAPI, commands, FAQ, documentation map Why it matters: This template simplifies the process of building AI SaaS applications, allowing developers and AI coding agents to focus on their unique features without the need for weeks of boilerplate code. The MIT-licensed MIT-licensed template is agent-optimized, making it easy for AI coding agents to contribute to the project. Backblaze B2 provides S3-compatible object storage at a fraction of typical cloud egress cost, ensuring generated media and uploads stay cheap.