AI Trending

Project directory

Projects

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

Java 9 stars

Amir2244/movies-rating

"movies-rating" is a recommendation system project that leverages distributed frameworks. Which includes services such as Hadoop Namenode, Hadoop Datanode, Spark Master, Spark Worker, and Redis.

### Overview: The "movies-rating" project is a comprehensive movie recommendation system that leverages distributed frameworks, including Hadoop, Spark, and Redis, to process user interactions, generate personalized recommendations, and provide analytics insights. ### Key Features: - **Microservices Architecture**: The system consists of multiple interconnected components that work together to provide a complete movie recommendation platform. - **Collaborative Filtering**: Uses user and item factors to generate personalized recommendations. - **Analytics Insights**: Provides interactive dashboards for visualizing data and generating insights. - **Real-Time Processing**: Processes user interaction events in real-time for immediate analysis and recommendation generation. - **Batch Processing**: Trains recommendation models using historical data and generates analytics insights. ### Tech Stack: Java ### Why It Matters: This project is essential for developers looking to build scalable and efficient movie recommendation systems. Its use of distributed frameworks and microservices architecture ensures that the system can handle large volumes of data and provide real-time recommendations, making it suitable for both small and large-scale applications.
Python 9 stars

AngeVox/AngeMedia-gateway

OpenAI-compatible media generation gateway for AI Agents. Supports multi-provider image/video generation, model routing, prompt enhancement, and built-in Web Studio.

### Overview: AngeMedia Gateway is an open-source media generation gateway for AI agents, designed to be self-hosted and compatible with OpenAI's models. It supports image and video generation, routing, and enhanced prompts, making it a versatile tool for developers and AI enthusiasts. ### Key Features: - **OpenAI-Compatible Image Generation**: Supports `POST /v1/images/generations`. - **Asynchronous Video Generation**: Handles `POST /v1/videos` with status check at `GET /v1/videos/{task_id}`. - **Cost-Aware Routing**: Routes across multiple providers including SiliconFlow Kolors, ModelScope, Pollinations, OpenAI, and explicit Agnes Image channels. - **Catalog-Driven Editing**: Enables image editing across predefined capabilities like SiliconFlow/Kolors, Qwen Image Edit, OpenAI GPT Image 2.5, Seedream 5, and Pollinations Edit. - **Stable Video Path**: Agnes Video v2.0 as the default, with `video_id` polling and `metadata.url` handling for Agnes Video 2.5. - **Offline Package**: fnOS/FYGO offline package for x86_64 and ARM64, including administrator credential recovery and database backup. - **Queue Execution**: Supports queue-first execution with durable SQLite job state, brokerless local backend for single-node installs, and optional Redis/Celery workers for higher concurrency. - **Protected Media Storage**: Local media import under `/generated/*` and `/uploads/*`. - **Web Studio**: Provides a user
Python 9 stars

Anil-matcha/wan-3.0-comfyui

ComfyUI custom nodes for Wan 3.0 — text-to-image, image edit, text-to-video, and image-to-video via MuAPI.

Overview: This ComfyUI custom node pack provides focused nodes for the Wan 3.0 API, allowing users to generate videos and images with specific features and controls. Key features: - Wan 3.0 Text-to-Video - Wan 3.0 Image-to-Video - Wan 3.0 Reference-to-Video - Wan 3.0 Spicy Text-to-Video - Wan 3.0 Spicy Image-to-Video - Wan 3.0 Spicy Reference-to-Video Tech stack: Python Why it matters: This pack simplifies the process of generating high-quality videos and images using the Wan 3.0 API, making it easier for developers to integrate advanced AI capabilities into their projects without needing extensive knowledge of MuAPI. Repository: Anil-matcha/wan-3.0-comfyui GitHub stars: 9 GitHub description: ComfyUI custom nodes for the Wan 3.0 API — text-to-image, image edit, text-to-video, and image-to-video via MuAPI.
Python 9 stars

BjornMelin/qdrant-neo4j-crawl4ai-mcp

MCP server combining Qdrant vector search, Neo4j knowledge graphs, and Crawl4AI web intelligence with agentic RAG capabilities. FastMCP 2.0 architecture with enterprise security, monitoring, and Kubernetes deployment. AI/ML engineering powerhouse.

Overview: This project is an Agentic RAG MCP Server that integrates Qdrant vector search, Neo4j knowledge graphs, and Crawl4AI web intelligence, providing autonomous orchestration capabilities. It combines vector intelligence, graph intelligence, and web intelligence to create a production-ready MCP server with enterprise security and monitoring features. Key features: - Qdrant vector search for semantic search - Neo4j knowledge graphs for knowledge representation - Crawl4AI web intelligence for smart web crawling and content extraction - Autonomous query routing and result fusion - Production-ready architecture with enterprise security and monitoring Tech stack: Python 3.11+, FastMCP 2.0, Qdrant, Neo4j, Crawl4AI, Docker Why it matters: This MCP server combines cutting-edge AI technologies to create a powerful, autonomous, and secure platform for intelligent coordination of multiple AI services, making it ideal for enterprise environments and AI/ML engineering projects.
Python 9 stars

Blackwellboy/laguna-s21-lab

Laguna S 2.1 (NVFP4) on DGX Spark GB10: 20-cell tuning sweep, 12h soak (3,099 turns, raw logs), thinking-gate dose-response study, Qwen 35B head-to-head, quant-floor reference. Losers published too.

### Overview: This repository documents the testing and analysis of the Laguna S 2.1 (NVFP4) model on a DGX Spark GB10, focusing on a 20-cell tuning sweep, 12-hour production soak, and a thinking-gate dose-response study. The raw logs and detailed results are provided to understand the model's performance and behavior. ### Key features: - **20-cell Tuning Sweep**: Detailed raw JSONs for every cell in the 20-cell tuning sweep. - **Container Recipe**: Pinpointed image and base digests, and wheel checksums for reproducibility. - **Benchmark Harness**: `hermes_bench_v1` used for throughput measurements. - **12-hour Production Soak**: Raw per-turn logs (3,099 turns), incident logs, integrity probes, and service memory samples. ### Tech stack: Python ### Why it matters: This repository provides a comprehensive and detailed look into the performance and behavior of the Laguna S 2.1 model, serving as a foundational resource for understanding and replicating the results.
Markdown 9 stars

Chaoses-Ib/VisualComputing

### Overview: Visual Computing is a comprehensive project that covers various aspects of image and video processing, computer vision, and image and video sharing. It includes features such as color spaces, image sensors, cameras, image scaling, data augmentation, diffusion models, image hosting, and sharing platforms like Pixiv and Danbooru. ### Key Features: - Color Spaces - Image Sensors - Cameras (Webcams, Surveillance Cameras) - Image Scaling - Data Augmentation - Diffusion Models (Latent Diffusion, Low-Rank Adaptation, ControlNet) - Image Hosting - Image Sharing (Pixiv, Danbooru) - Video Processing (Video File Formats, Subtitles, Video Players, Video Scaling, Video Enhancement, Frame Interpolation) - Computer Vision (Features, Image Similarity, Image Classification, Image Segmentation, Object Detection, Object Recognition (YOLO, Optical Character Recognition, Face Recognition)) - 3D (Pose Estimation, 3D Reconstruction, Object Tracking) - Libraries (Open Source Computer Vision Library) ### Tech Stack: Markdown ### Why It Matters: Visual Computing provides a broad range of tools and techniques for processing and analyzing images and videos, making it essential for various applications in fields such as photography, computer vision, and multimedia sharing.
Jupyter Notebook 9 stars

Chirayu-Tripathi/VerAIzon

RAG (Retrieval Augmented Generation) chatbot accompanied with Mistral-7B, specifically tailored for Verizon customer services.

### Overview: VerAIzon is a RAG (Retrieval Augmented Generation) chatbot designed for Verizon customer services, utilizing Mistral-7B for enhanced performance. The chatbot processes large datasets and efficiently retrieves relevant information from a vector database, providing personalized and accurate responses to customer queries. ### Key Features: - **Vector Database Storage**: Data is stored in a FAISS vector database for efficient similarity search. - **Prompt Generation**: Custom prompts are generated using the most relevant context chunks from the vector database. - **Retrieval Augmentation**: Mistral-7B is used to enhance the chatbot's ability to provide contextually relevant answers. - **Data Extraction**: The chatbot uses iterative data extraction from web pages and user guides to build a comprehensive dataset. ### Tech Stack: - **Programming Languages**: Jupyter Notebook - **Vector Database**: FAISS (Facebook AI Similarity Search) - **Model**: Mistral-7B (Open-Source State of the Art) ### Why It Matters: VerAIzon streamlines customer service by leveraging advanced AI models and efficient data storage techniques, ensuring faster and more accurate responses to customer inquiries.
Rust 9 stars

Cognilogical/NeuroStrata

Cognitive Memory Architecture for Agents

### Overview: NeuroStrata is a cognitive memory architecture designed for AI agents, providing six distinct ways to think, remember, plan, and verify work. It eliminates the trade-off between flat markdown files and manual configuration, offering a single compiled Rust binary that automates the entire process. ### Key Features: - **Tri-Strata Memory:** Persistent Global, Project Domain, and Task Working Memory. - **Negative Exemplars:** Failed approaches are stored as evidence, preventing amnesia. - **Procedural Memory:** Recurring routines fire as tracked procedures. - **Live Software Structure:** Incremental Tree-sitter parsing and automatic graph-first reflex. - **Deductive & Semantic Thinking:** Native W3C Semantic Web engine for exact logic. ### Tech Stack: Rust ### Why It Matters: NeuroStrata simplifies the complex process of developing AI agents by providing a comprehensive and automated solution that eliminates the need for manual configuration and trade-offs between flat markdown files and complex systems.
HTML 9 stars

DarrylBayliss/Simon-Says-RAG-Android

An Android App recreating the Simon Says game. Uses MediaPipe to run an LLM on device using RAG

Overview: An Android app that recreates the Simon Says game, utilizing MediaPipe to run a Gemma3 1b LLM and a RAG pipeline on the device. Key features: - Simon Says game implementation - MediaPipe integration for game logic - RAG-powered language model for app functionality Tech stack: HTML Why it matters: This project showcases the integration of AI and media processing on Android devices, demonstrating how to run a Gemma3 1b LLM and a RAG pipeline on the device, providing a unique gaming experience. Repository: DarrylBayliss/Simon-Says-RAG-Android Primary language: HTML GitHub stars: 8 GitHub description: An Android App recreating the Simon Says game. Uses MediaPipe to run the Gemma3 1b LLM and a RAG pipeline on device. 📣 🦾
TypeScript 9 stars

DavidGir/CamelotAI

Camelot AI lets you upload documents, chat about their content, and get audible responses from the chatbot. It's perfect for digging into details or just getting a quick summary—all with text-to-speech for easier access and a better experience. Explore Camelot AI and make your documents talk!

Overview: Camelot AI allows users to interact with their documents by uploading them, chatting about their content, and receiving audible responses. It simplifies the process of exploring details or getting quick summaries, with text-to-speech capabilities for a better user experience. Key features: - Upload documents - Chat about document content - Get audible responses Tech stack: Next.js 14 App Router, GPT-4 through OpenAI, all-MiniLM-L6-v2 through HuggingFace, LangChain.js, ElevenLabs, Pinecone, Bytescale, Vercel, Clerk, Tailwind CSS. Why it matters: Camelot AI streamlines the interaction with documents, making them more accessible and engaging through text-to-speech, providing a unique and convenient way to explore information and insights.
TypeScript 9 stars

Dogwind221/img2img-studio

Image-to-image studio: vision-gated routing + multi-provider generation (DashScope / ChatGPT-web / OpenAI / Z.AI / Doubao / MiniMax) with auto failover. E-commerce sets & photo-art styles. 图生图工作台 · 多供应商自动降级 · 电商套图与照片艺术

### Overview: Image-to-image studio is a versatile image transformation tool designed for developers working with AI agents. It enables users to quickly generate various image styles and designs by selecting from multiple providers, including DashScope, ChatGPT-web, OpenAI, Z.AI, Doubao, and MiniMax. The tool supports both e-commerce sets and photo-art styles, making it a valuable asset in the field of AI-driven design and development. ### Key features: - **Multi-Provider Generation**: Supports multiple image generation providers including DashScope, ChatGPT-web, OpenAI, Z.AI, Doubao, and MiniMax. - **Auto Failover**: Automatically switches to a backup provider in case of failure. - **Auto Model Chain**: Routes the image generation through a chain of models to ensure the best possible quality. - **E-commerce Sets**: Generates sets of images suitable for e-commerce platforms, including white background product images, detail shots, and more. - **Photo-Style Art**: Creates editorial, revival, fusion, and zine-style images. - **ChatGPT Web Outfit**: Allows generating images without an API key using ChatGPT Web accounts. - **Multi-Provider Auto-Downgrade**: Supports multiple providers for model downgrade, including DashScope, OpenAI, Z.AI, MiniMax, and local ChatGPT Web services. - **Batch Generation**: Supports batch processing for e-commerce sets, allowing for efficient image creation. - **Quality Pre-Settings**: Offers quality presets for 1k and 2k images, and supports i2i reference images
Python 9 stars

Eliezer318/SPOT-dLLM

SPOT: lossless speculative decoding for diffusion language models via commit-order prediction

### Overview: SPOT: lossless speculative decoding for diffusion language models via commit-order prediction. This project aims to enhance speculative decoding for diffusion language models by predicting the commit order and verifying drafts in one pass. The project has been published on the arXiv and is open-source, with a draft available on Hugging Face. ### Key features: - **SPOT**: A 0.62M-parameter drafter that predicts the commit order and verifies drafts in one pass. - **SPOT-γ**: A variant that drafts whole commit sets and raises SDAR-8B's throughput by 59% over the threshold decoder alone. - **SPOT-Sets**: A dynamic threshold decoding approach that drafts whole commit sets and improves throughput by 59% over the threshold decoder alone. ### Tech stack: Python ### Why it matters: SPOT speeds up diffusion language models losslessly, achieving up to 3.4 times the throughput of standard decoding. It also transfers drafters across six diffusion LMs, making it a valuable tool for researchers and developers working with diffusion models.
Python 9 stars

Enntity/sparkscore

Typed, calibrated decisions on one NVIDIA DGX Spark: d3-ultra 27B NVFP4 on Atlas, JevBench public 90.9% at p50 146 ms, every question in one engine pass.

### Overview: SparkScore is a powerful tool for making typed, calibrated decisions on a single NVIDIA DGX Spark. It leverages the d3-ultra model quantized to NVFP4, providing high accuracy and low latency. ### Key Features: - **Typed Decisions**: SparkScore provides yes/no, choice, and ordinal level answers. - **Fast Execution**: It runs on the DGX Spark, delivering results in milliseconds. - **High Accuracy**: Achieves 90.9% accuracy on JevBench v1.6.1, with a latency of 146 ms. - **4.3x Faster**: Compared to d3's own runtime, SparkScore is 4.3x faster, with a latency of 1.05 seconds. ### Tech Stack: Python ### Why It Matters: - **High Accuracy and Low Latency**: SparkScore excels in providing accurate and fast decision-making capabilities. - **Versatile Model**: Utilizes the d3-ultra model, quantized to NVFP4, offering high performance and reliability. - **Scalability**: Runs on the DGX Spark, ensuring data remains on the hardware, maintaining privacy and security. ### Repository: Enntity/sparkscore - **Stars**: 8 - **Description**: Typed, calibrated decisions on one NVIDIA DGX Spark: d3-ultra 27B NVFP4 on Atlas, JevBench public 90.9% at p50 146 ms, every question in one engine pass.
Python 9 stars

Eyna-A/STAIR-ToC-RAG

An open-source implementation of ToC-Aware RAG inspired by IBM's STAIR paper. Preserves document hierarchy for structured addressable retrieval and near-zero hallucinations.

Overview: STAIR-ToC-RAG is an open-source implementation of a document-native retrieval framework inspired by the IBM Research paper "STAIR: A novel dataset and LLM based retriever for document structure augmentation." It preserves the document hierarchy for structured addressable retrieval and mitigates near-zero hallucinations. Key features: - ToC Tree Extraction - Hierarchical Address Injection - ToC-Aware Retriever - Pluggable Vector Store (Qdrant by default) - REST API - Interactive Dashboard - Container-first Architecture Tech stack: Python Why it matters: - Fewer hallucinations - No lost context - Auditable answers - Built for long, structured documents GitHub stars: 8 Repository: Eyna-A/STAIR-ToC-RAG
Python 9 stars

FluffyAIcode/LLM-KV--Cache-compress

Discrete Kakeya cover for LLM KV cache: D4/E8 nested-lattice quantisation realising a Kakeya-style tube-cover over the direction sphere. 2.4x-2.8x compression at <1% perplexity loss on Qwen3, Llama-3, DeepSeek, GLM-4, Gemma. Drop-in transformers.DynamicCache. pip install kakeyalattice.

### Overview: Discrete Kakeya Cover for LLM KV Cache Compression ### Key Features: - **KakeyaLatticePackedCache** stores bit-packed lattice codes, achieving ~2.4x HBM compression for Q=38. - **KakeyaLatticeQuantizedCache** stores int8 lattice indices, offering ~1.94x HBM compression. - **KakeyaLatticeCache** reconstructs bf16, providing zero HBM savings. ### Tech Stack: Python ### Why It Matters: At the codec bit-rate level, this project achieves 2.4x–2.8x compression at <1% perplexity loss on Qwen3, Llama-3, DeepSeek, GLM-4, and Gemma, demonstrating significant improvements over existing quantization methods.
Python 9 stars

GoodbyePlanet/semcode

Semantic code-search (semcode) MCP. Indexes code symbols and commit history. Combines dense embeddings with sparse BM25 vectors for hybrid search that balances semantic understanding with keyword precision.

Overview: semcode is a powerful semantic code-search tool that indexes code symbols and commit history across multiple GitHub repositories, providing a hybrid search solution that combines dense embeddings with sparse BM25 vectors for improved semantic understanding and keyword precision. Key features: - Semantic code search across multiple repositories - Indexes code symbols and commit history - Combines dense embeddings with sparse BM25 vectors for hybrid search - Serves as an MCP server for AI agents - Supports Python 3.12+ and CI Tech stack: Python Why it matters: semcode is essential for coding agents to efficiently search and understand code across different repositories, providing a more comprehensive and context-aware search experience compared to traditional keyword-based methods.
Python 9 stars

Green0-0/light_prompter

Accelerate test-time-compute with batching!

### Overview: Light Prompter is a Python framework designed to accelerate test-time-compute by efficiently batching responses with prompting strategies that include multiple steps. This approach allows for significant speedups when using inference engines like vLLM. ### Key Features: - **Batching Responses**: Light Prompter aggregates chat completions into batches, enabling faster processing. - **Multiple Step Strategies**: It supports various prompting strategies such as self-consistency, two-step response, and more. - **State Machines**: Responders act as state machines, each containing nested responders that request chat completions. ### Tech Stack: - Python ### Why It Matters: Light Prompter offers substantial performance improvements by batching responses, making it an essential tool for developers working with inference engines and test-time-compute tasks. ### Repository: Repository: Green0-0/light_prompter GitHub Stars: 9 GitHub Description: Accelerate test-time-compute with batching!
Python 9 stars

HodeTech/ForgeLM

Config-driven LLM fine-tuning with safety evaluation, EU AI Act compliance, 6 alignment methods, and one-command bundled quickstart templates.

### Overview: ForgeLM is a powerful tool for fine-tuning Large Language Models (LLMs) with safety and compliance features. It simplifies the process of fine-tuning models, ensuring reproducibility, data integrity, and compliance with regulations like the EU AI Act. ### Key Features: - **Validated YAML Configuration**: ForgeLM uses validated YAML configurations to ensure model consistency and reproducibility. - **Fine-Tuning and Evaluation Gates**: It supports fine-tuning with preference and reasoning methods, and includes evaluation gates for safety and compliance. - **Data Audit and Quality Checks**: ForgeLM performs dataset audits to ensure data quality and compliance with regulations. - **Conditional Promotion**: It allows for conditional promotion of models based on loss, benchmarking, safety, and human approval. - **Reviewable Evidence**: ForgeLM provides hash-chained audit logs, model integrity manifests, and model cards for transparency and traceability. - **Optional Compliance Export**: It supports optional compliance exports for regulatory compliance. - **Support for Multiple Models and Pipelines**: ForgeLM supports fine-tuning existing Hugging Face models and can be used in pipelines like SFT → DPO → GRPO. - **Use of Experimental APIs**: ORPO and SimPO may use experimental TRL APIs, but these are optional. - **Integration with Other Tools**: ForgeLM integrates with tools like Ollama, vLLM, TGI, or llama.cpp for serving. ### Tech Stack: Python ForgeLM is built using Python, leveraging its robust ecosystem for machine learning and automation. ### Why It Matters: Forge
Python 9 stars

Infatoshi/traces

Extract, redact, and export AI coding-agent sessions into privacy-safe training JSONL

### Overview: `traces` is a Python tool designed to extract, redact, and export AI coding-agent sessions into privacy-safe JSONL format. It normalizes multi-harness session stores into a conversational schema, redacts secrets and paths, and exports data in DistillKit-compatible formats. ### Key Features: - Normalizes multi-harness session stores into a conversational schema. - Redacts secrets, paths, emails, and identifiers to stable placeholders. - Exports data in DistillKit-compatible chat formats and causal agent-step formats. - Verifies the integrity of the exported data before publishing. ### Tech Stack: Python ≥ 3.10 ### Why it Matters: Raw agent logs are not training data. Formats differ by product, and sessions contain sensitive information. This tool normalizes, redacts, and exports data in a privacy-first manner, ensuring that only necessary information is retained and the data is ready for training purposes. ### Repository: Repository: Infatoshi/traces Primary Language: Python GitHub Stars: 9 GitHub Description: Extract, redact, and export AI coding-agent sessions into privacy-safe JSONL GitHub URL: [Infatoshi/traces](https://github.com/Infatoshi/traces) License: MIT — extractor/redaction provenance in [ATTRIBUTION.md](./ATTRIBUTION.md) ### Related Projects: - [0xSero/ai-data-extraction](https://github.com/0xSero/ai-data-extraction) - Source extractors this repo packages and extends -
Unknown 9 stars

JimPresting/supabase-oracle-cloud-docker

Complete guide to self-host Supabase on Oracle Cloud free tier with Docker, Nginx, SSL, and automatic updates. Includes auto-start after reboots and optimized backup system.

Overview: This guide provides a step-by-step process for setting up a Supabase instance on Oracle Cloud using Docker, Nginx, SSL certificates, and automatic updates. It includes auto-start after reboots and an optimized backup system. Key features: - Docker setup - Nginx integration - SSL certificate management - Automatic updates - Auto-start after reboots - Optimized backup system Tech stack: Docker, Nginx, SSL certificates, Oracle Cloud, Supabase Why it matters: This guide simplifies the process of self-hosting Supabase, making it accessible for developers without requiring extensive technical knowledge.