Project directory

Projects

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

Jupyter Notebook 18 stars

bentoml/Fraud-Detection-Model-Serving

Online model serving with Fraud Detection model trained with XGBoost on IEEE-CIS dataset

Overview: This project demonstrates how to deploy a fraud detection model using BentoML, a Python library for building and deploying machine learning models. The model is trained on the IEEE-CIS dataset using XGBoost and is available for online serving. Key features: - Deployed model using BentoML - Trained on the IEEE-CIS Fraud Detection competition dataset - Available for online serving - Model variations: tiny, large, small Tech stack: BentoML, XGBoost, Jupyter Notebook Why it matters: This project showcases the ease of deploying machine learning models using BentoML, a powerful tool for model serving that simplifies the process of creating, training, and deploying models. The use of BentoML allows for efficient model management and deployment, making it a valuable resource for developers looking to integrate machine learning into their projects.

Python 18 stars

brianSalk/JSONLgenerator

an easy way to create JSONL files for fine-tuning openai models.

Overview: A tool designed to simplify the creation of JSONL files for fine-tuning OpenAI models, making the process more accessible and efficient. Key features: - Generates JSONL files for fine-tuning GPT3.5 and newer models. - Streamlined command-line interface for easy execution. - Web application available for direct integration. Tech stack: Python Why it matters: This project offers a straightforward solution for developers looking to fine-tune models, making the process more user-friendly and accessible. Repository: brianSalk/JSONLgenerator GitHub stars: 18

Unknown 18 stars

casper-vdb/Casper

⚡ World’s Fastest Vector Database for AI & RAG

### Overview: Casper is a high-performance vector database designed for AI and RAG applications, offering exceptional performance and scalability. ### Key Features: - **Rust Backend**: Built with Rust for optimal performance and reliability. - **Python, Go, Rust Clients**: Available for easy integration. - **Scalability**: Supports high load and real-time search systems. - **Performance**: Consistently outperforms Qdrant across various benchmarks, delivering up to an order of magnitude higher RPS. ### Tech Stack: Rust 🦀 ### Why Casper? - **Performance Leader**: Casper consistently outperforms Qdrant in our benchmarks, achieving up to an order of magnitude higher RPS. - **Infrastructure Savings**: Requires fewer compute resources, leading to lower costs and more efficient use of resources. - **Real-time Search & AI**: Ideal for AI and RAG applications, providing fast and scalable vector search capabilities. ### Casper vs Qdrant - **Benchmarks**: - **RPS & Recall**: - **Casper**: 248.86 k Top@10, 93.55 k Top@100, 12.02 k Top@1k, 1.43 k Top@10k, 163 Top@100k - **Qdrant**: 13.25 k Top@10, 15.65 k Top@100, 2.48 k Top@

Jupyter Notebook 18 stars

cilabuniba/i-dream-my-painting

[WACV 2025] I Dream My Painting: Connecting MLLMs and Diffusion Models via Prompt Generation for Text-Guided Multi-Mask Inpainting

### Overview: This repository contains the reference code for the paper "I Dream My Painting: Connecting MLLMs and Diffusion Models via Prompt Generation for Text-Guided Multi-Mask Inpainting" presented at WACV 2025. The project aims to connect Multi-Modal Language Models (LLMs) and Diffusion Models to generate text-guided multi-mask inpaintings. ### Key features: - Multi-Mask Inpainting with RCA (Reconstructive Contour Augmentation) - Text-Guided Inpainting using Prompt Generation - Integration of MLLMs and Diffusion Models for Image Restoration - Stable Diffusion 2.1 checkpoint for model weights ### Tech stack: Jupyter Notebook ### Why it matters: This project demonstrates the potential of combining MLLMs and Diffusion Models to achieve high-quality text-guided image inpaintings, which can be useful in various applications such as art restoration, image editing, and generative design.

JavaScript 18 stars

civitai/civitai-gen-skill

Generate images, video, audio and more via Civitai's orchestration API — a runtime-agnostic agent skill (AgentSkills format, zero deps, Node 18+).

### Overview: Civitai's orchestration API provides a runtime-agnostic agent skill that allows developers to generate various media types such as images, videos, audio, and more. This skill is designed to work with any skills-compatible agent runtime, including Claude Code, OpenClaw, and others. It is a zero-dependency Node.js CLI that supports zero-buzz generation and is compatible with the Civitai MCP server for model discovery and signed-in write actions. ### Key Features: - **Images**: Supports text-to-image, img2img, checkpoint/LoRA selection, and resolution control. - **Video**: Handles 11+ engines, including Veo 3, WAN, Kling, LTX2, Hunyuan, and more, with optional audio. - **Text-to-speech**: Built-in speakers and voice cloning capabilities. - **Music**: ACE-Step song generation with lyrics and duration control. - **Transcription**: Speech-to-text with timestamps and language hints. - **Bulk & experiments**: Supports batch processing and wildcard expansion. - **Cost estimation**: Provides dry-run buzz estimates for generation. ### Tech Stack: - **Node.js 18+**: Uses native `fetch` for efficient handling of media generation. - **Agent Skills Format (AgentSkills)**: Supports zero dependencies and runs on Node.js. ### Why It Matters: This skill streamlines the process of generating media content using Civitai's orchestration API, making it accessible and efficient for developers. It simplifies the integration of media generation into existing agent skill ecosystems, enabling a

TypeScript 18 stars

devyanshyadav/librenote-ai

The open-source, self-hosted alternative to NotebookLM.

### Overview: LibreNote AI is an open-source, self-hosted alternative to NotebookLM, providing users with the freedom to control their data and choose from a variety of AI models, including any model available on OpenRouter. It offers a wide range of studio artifacts such as mind maps, flashcards, quizzes, reports, and diagrams, and allows for the generation of audio overviews. The platform is built on PostgreSQL and Supabase, ensuring data privacy and user control over their uploads. ### Key Features: - **Self-Hosted**: Users can host the app on their own machines or servers. - **Data Privacy**: Data is processed locally, not in the cloud. - **AI Models**: Supports models from OpenRouter, including Gemini, Llama, and more. - **Studio Artifacts**: Offers 8 types of studio artifacts like mind maps, flashcards, quizzes, reports, data tables, diagrams, audio, and notes. - **Open Source**: MIT License, allowing for inspection, fork, and extension. - **Multi-User**: Supports multi-user authentication with per-user data isolation. - **Open Source**: Built on PostgreSQL and Supabase, ensuring data privacy and user control. ### Tech Stack: - **Backend**: Built with TypeScript, using the Next.js framework. - **Frontend**: Developed with React, leveraging the Supabase API for database operations and Supabase UI for user interface. - **Database**: PostgreSQL for storing user data and project information. - **Storage**: Supabase for managing user uploads and project files. ### Why It

Python 18 stars

digitalspaceport/Multi-Agent-Benchmark-Tool

Interested in running a conspiracy of agents (technical term) on your Local AI Infra? Who isn't!

### Overview: This lightweight benchmark tool is designed to simulate multiple autonomous agents interacting with a single OpenAI-compatible endpoint, focusing on request-level metrics such as Time-To-First-Token (TTFT), Time-Per-Output-Token (TPOT), and throughput. It is particularly useful for tuning inference stacks in a local AI infrastructure. ### Key features: - **TTFT (Time-To-First-Token)**: Measures the time from request start to the first streamed response event. - **TPOT (Time Per Output Token)**: Calculates the completion latency divided by completion tokens, providing insights into text-generating responses. - **Prefill Throughput**: Measures the rate of prefill tokens per second. - **Decode Throughput**: Measures the rate of completion tokens per second. - **Requests per Second (RPS)**: Measures the number of completed requests per second. - **Realistic Agent Behaviors**: Includes tool calls with streamed aggregation, follow-up assistant requests, and multimodal user inputs. - **Robustness**: Supports usage-based token counts during streaming and fallback token estimation when metadata is unavailable. - **Pacing Controls**: Allows for controlled pacing with jitter to avoid synchronized spikes and timeouts and SDK-level retries. ### Tech stack: Python ### Why it matters: - **Request-level Metrics**: Provides detailed metrics for agents, including TTFT, TPOT, and throughput, which are crucial for understanding the performance of agents. - **Realistic Behaviors**: Simulates agent interactions, including tool calls, follow-up requests, and multimodal inputs, which are essential for realistic

Jupyter Notebook 18 stars

dvianna/LegalQA-bloomz-560m

Finetuning a small BLOOMZ model (bloomz-560m) on a small dataset and with limited resources.

Overview: This repository finetunes a small BLOOMZ model (bloomz-560m) on a small dataset and with limited resources, aiming to create a legal question answering model called LegalQA-bloom-560. Key features: - LegalQA dataset built from Brazilian Internal Revenue Service pdfs - Prompt dataset generated for natural language instructions - Finetuned BLOOMZ model on prompts from prompts.json Tech stack: Jupyter Notebook Why it matters: This project demonstrates the feasibility of finetuning a large language model on a small dataset and limited resources, paving the way for more efficient legal AI solutions.

Python 18 stars

gchochla/VAuLT

This repo contains the original implementation of VAuLT, the Vision-and-Augmented-Language Transformer. We provide instructions to download some multimodal social-media datasets, and scripts to experiment with. VAuLT is a stack of Transformers, a LM like BERT that preprocesses the text input of ViLT

Overview: This repository contains the original implementation of VAuLT, a Vision-and-Augmented-Language Transformer designed for sentiment classification on social media. It includes scripts and instructions for downloading multimodal social-media datasets and experimenting with the model. Key features: - Scripts for downloading and preprocessing datasets such as Twitter-TMSC, MVSA, and Bloomberg. - Instructions for using VAuLT with pre-trained models and entity linking for Twitter datasets. - A stack of Transformers, including a Vision-and-Augmented-Language Transformer (VAuLT) model. Tech stack: Python Why it matters: VAuLT is a powerful tool for sentiment analysis on social media, leveraging the strengths of both vision and language models to provide more accurate and context-aware insights.

Java 18 stars

h2oai/mlops-dai-runtimes

Production ready templates for deploying Driverless AI (DAI) scorers. https://h2oai.github.io/dai-deployment-templates/

Overview: This repository contains production-ready templates for deploying Driverless AI (DAI) scorers, providing a structured approach for integrating and deploying DAI models. Key features: - Deployment templates for various target environments - Shared code and libraries for common functionalities - Gradle-based multi-project repository for easy distribution and integration - Integration with Driverless AI (DAI) scorers Tech stack: Java Why it matters: These templates simplify the deployment process, ensuring that DAI models are efficiently and reliably deployed across different environments, making them ideal for production use cases. Repository: h2oai/mlops-dai-runtimes Primary language: Java GitHub stars: 18 GitHub description: Production ready templates for deploying Driverless AI (DAI) scorers.

Python 18 stars

henrygouk/mars-finetuning

Overview: This repository contains the code used for conducting experiments in the paper "Distance-Based Regularisation of Deep Networks for Fine-Tuning," published at ICLR 2021. The code uses the keras API bundled with tensorflow and has been tested with the official TF 2.1 docker image. The `scripts/` directory contains bash scripts for reproducing the experiments in the paper. Key features: - Fine-tuning with MARS Distance Constraints - Theoretical and empirical evidence for improved generalisation performance - A distance metric based on the Maximum Absolute Row Sum (MARS) norm - Projected gradient descent as a regularisation strategy - A hard constraint on the MARS distance between pre-trained and fine-tuned weights in each layer Tech stack: Python Why it matters: This method can significantly improve the performance of fine-tuned networks by limiting the weight changes during the fine-tuning process, leading to better generalisation and potentially higher accuracy.

TypeScript 18 stars

hoangsonww/FRED-Data-Analysis

💸 A comprehensive AI-powered data explorer that combines FRED economic data & insights with vector search, regression analysis, and interactive RAG chatbot via Pinecone Vector DB, OpenAI, Claude, and Gemini. Built with TypeScript, React, and Express for seamless full-stack performance.

### Overview: A comprehensive AI-powered data explorer that combines FRED economic data and insights with vector search, regression analysis, and an interactive RAG chatbot using Pinecone Vector DB, OpenAI, Claude, and Gemini. Built with TypeScript, React, and Express for seamless full-stack performance. ### Key Features: - Fetches and vectorizes FRED economic data using Pinecone. - Provides a chatbot interface for querying the data. - Includes regression analysis on various economic indicators like linear, polynomial, daily percent change, and logarithmic regression. - Results are summarized in a report format, including charts and AI-generated insights backed by data. ### Tech Stack: - TypeScript - React - Express ### Why It Matters: This project offers a powerful tool for data-driven economic analysis, combining cutting-edge AI with traditional economic data exploration. It simplifies complex economic data analysis, making it accessible and insightful for developers and researchers alike.

Python 18 stars

i4Ds/whisper-prep

Data preparation utility for the finetuning of OpenAI's Whisper model.

### Overview: Whisper-prep is a data preparation utility designed for fine-tuning OpenAI's Whisper model. It assists in generating training data by synthesizing long-form audio from sentence-level clips, processing existing SRTs, and normalizing them to Netflix-style guidelines. ### Key Features: - **Sentence Concatenation for Long-Form Audio**: Combines multiple sentence-level audio clips into longer recordings with matching SRT subtitles, simulating real-world long-form transcription scenarios. - **SRT Processing and Audio Segmentation**: Converts existing long-form audio with SRT/VTT transcripts into Whisper-compatible 30-second chunks, preserving timestamp information and handling overlapping utterances. - **Netflix-Style SRT Normalization**: Merges short captions to meet Netflix subtitle guidelines (max 42 characters, max 7 seconds), removing problematic samples. ### Tech Stack: Python ### Why It Matters: Whisper-prep streamlines the data preparation process for fine-tuning Whisper, making it easier to train models on diverse and high-quality audio data. This utility is essential for researchers and developers looking to improve the performance of Whisper models on various tasks.

Java 18 stars

integrallis/vectors

The Ultimate Vectors Library for Java

Overview: The `vectors` library is a comprehensive Java-based vector search and persistence solution designed for JVM applications. It provides indexed similarity search, durable local state, and metadata filtering capabilities, built directly into the application without the need for external vector databases. Key features: - Built on the JDK Vector API for SIMD distance kernels - Offers a full menu of ANN indexes, quantizers, and a generation-based mmap storage engine - Provides Spring AI and LangChain4j adapters - Supports HNSW, Vamana, IVF-Flat, and IVF-PQ indexes - Includes eight quantizers - Atomic generation-based persistence - Built-in Spring AI-first collection API Tech stack: Java 25 Why it matters: The `vectors` library simplifies vector search and persistence in Java applications, offering a seamless integration that enhances performance and reliability without the need for separate services or external databases. This makes it ideal for applications that require efficient vector search capabilities and durable local state management. Repository: integrallis/vectors Primary language: Java GitHub stars: 18 GitHub description: The Ultimate Vectors Library for Java

Python 18 stars

j6qn3i24az-9a/seedream-4-0-api-cn

即梦图片 — Seedream 4.0 API (`seedream-4-0`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: 即梦图片是ByteDance开发的统一文本到图像和图像编辑模型，支持4K输出、行业领先的文本渲染和批量处理功能。 Key features: - 一个Key，整个目录。同一个`https://api.apimart.ai/v1` base URL和`Authorization`请求头即可访问Seedream 4.0（`seedream-4-0`）以及图像、视频、语言等300+模型——只需切换`model`字段，客户端无需改动。 - $1起充，按量付费。没有订阅、也没有需要预先估算用量的预付费套餐：最低充值$1，调用多少扣多少。 - 费用随响应一并返回。每个完成的任务都会报告实际计费金额（`cost` / `credits_cost`）；状态为`failed`的任务会自动退款。 - 原生异步设计。提交任务、拿到`task_id`、轮询`GET /v1/tasks/{id}`——批处理和重试都是普通的队列工作，不需要定制集成。 Tech stack: Python Why it matters: 即梦图片是ByteDance开发的统一文本到图像和图像编辑模型，支持4K输出、行业领先的文本渲染和批量处理功能。它提供了一个OpenAI兼容的端点`https://api.apimart.ai/v1`，无需包月套餐，最低充值$1，按量付费。

Jupyter Notebook 18 stars

juletx/spatial-reasoning

Grounding Language Models for Compositional and Spatial Reasoning

Overview: This project aims to evaluate grounded neural language models for compositional and spatial reasoning, addressing the limitations of current vision-and-language models. Key features: - State-of-the-art pre-trained and fine-tuned vision-and-language models evaluation - Development of synthetic datasets focusing on compositional reasoning - Improvements in compositional reasoning and zero-shot spatial reasoning experiments Tech stack: Jupyter Notebook Why it matters: This project addresses the challenges in grounding language models for compositional and spatial reasoning, providing a valuable contribution to the field of artificial intelligence and natural language processing.

Python 18 stars

k7v530p627/wan-2-7-videoedit-api-id

Wan 2.7 Video Edit API (`wan2.7-videoedit`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Wan 2.7 Video Edit API (`wan2.7-videoedit`) is an advanced AI video editing service that offers superior visual quality, high motion coherence, and professional-grade performance. It supports multiple video and image formats and allows for optional reference images for the target style or appearance. Key features: - Edit an existing source video - Use optional instructions - Support up to four reference images Tech stack: Python Why it matters: This API provides a highly efficient and cost-effective solution for video editing, allowing for quick and easy integration into various projects without the need for a monthly subscription. The API supports multiple resolutions and formats, ensuring compatibility with different devices and platforms. Repository: k7v530p627/wan-2-7-videoedit-api-id Primary language: Python GitHub stars: 10 GitHub description: Wan 2.7 Video Edit API (`wan2.7-videoedit`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Rust 18 stars

kn0sys/valentinus

A thread-safe vector database for model inference inside LMDB.

Overview: Valentinus is a thread-safe vector database designed for model inference within LMDB, optimized for performance and ease of use. Key features: - Thread-safe vector database implementation - Integration with LMDB for efficient storage and retrieval - Support for model inference and embeddings Tech stack: Rust Why it matters: Valentinus simplifies the process of integrating vector databases into existing models, making it easier to manage and utilize vector data in applications. Repository: kn0sys/valentinus GitHub stars: 18 GitHub description: A thread-safe vector database for model inference inside LMDB.

Python 18 stars

kr4ckhe4d/local-llm-benchmarks

Measured llama.cpp benchmarks on AMD Radeon RDNA4 with ROCm: RX 9070 XT + Radeon AI PRO R9700 (48 GB). Qwen3.8, Gemma 4, GLM, gpt-oss; GGUF quant quality (KLD), MTP speculative decoding, 128K context, Claude Code with local models.

### Overview: This repository measures and benchmarks local language models (LLMs) on AMD Radeon RDNA4 GPUs using ROCm, focusing on performance metrics such as generation speed, VRAM usage, and quantization quality. ### Key Features: - **GPU Performance**: Measured on an RX 9070 XT (16 GB) and a Radeon AI PRO R9700 (32 GB), both RDNA4 / gfx1201, with ROCm 7.2 under Linux. - **Models Tested**: Qwen3.8, Gemma 4, GLM, gpt-oss; GGUF quant quality (KLD), MTP speculative decoding, 128K context, Claude Code with local models. - **Tech Stack**: Shell scripts and raw logs are available in the `benchmarks/` directory. ### Why It Matters: This repository provides detailed measurements and benchmarks of local LLMs on consumer AMD GPUs, offering insights into performance and optimization techniques for developers working with these models.

Shell 18 stars

kylemclaren/qdrant-on-fly

Bootstrap a Qdrant vector database cluster on Fly.io

Overview: This repository provides a highly available Qdrant vector database cluster on Fly.io, suitable for running on a private network with peer-to-peer communication and discovery. Key features: - Qdrant vector database cluster - Private network setup - P2P communication and discovery - Fast and reliable performance with Rust Tech stack: Rust Why it matters: This setup allows for efficient and scalable vector database management, particularly useful for applications requiring advanced filtering capabilities and integration with modern LLMs. GitHub stars: 18