Project directory

Projects

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

Python 12 stars

ke84b1y8yw17/gemini-3-1-flash-lite-image-prompts

Nano Banana 2 Lite prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-3.1-flash-lite-image-ext`) and the cost per image.

Overview: This repository contains Nano Banana 2 Lite prompts designed for image rendering, including 8 tested recipes with example renders, aspect ratios, the model ID (`gemini-3.1-flash-lite-image-ext`), and the cost per image. Key features: - 8 tested recipes - Example renders - Aspect ratios - Model ID (`gemini-3.1-flash-lite-image-ext`) - Cost per image Tech stack: Python Why it matters: This repository provides a practical guide for developers looking to use Nano Banana 2 Lite with OpenAI-compatible endpoints, offering cost-effective and efficient image prompts without the need for a monthly subscription.

Python 12 stars

kizlin5j9/doubao-seedance-1-5-pro-prompts-ja

シードダンス / シーダンス — Seedance 1.5 Pro prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`seedance-1.5-pro`) and the cost per image.

### Overview: Seedance 1.5 Pro Prompts: 8 Tested Recipes With Outputs  <p align="center"> <img src="assets/badges/price.svg" alt="observed unit price"> <img src="assets/badges/billing.svg" alt="billing model"> <img src="assets/badges/compat.svg" alt="OpenAI-compatible endpoint"> </p> <p align="center"> <img src="assets/01-product-ad-ceramic-mug-thumb.jpg" width="820" alt="Sample video frame generated through APIMart"> </p> > **from $0.0104 per second** on the APIMart route (billed by resolution tier) — `https://api.apimart.ai/v1` の OpenAI 互換エンドポイント1つだけ、月額プランは不要です。*(2026-10-12 時点)* **[Get an API key](https://go.apimart.ai/k-6d966b)** · **[Live pricing](https://go.apimart.ai/k-d3d2c9)** · **[Model page](https://go.apimart.ai/k-108683)** · [⚡ 60-second quickstart](#quickstart) **チームが APIMart 経由で Seedance 1.5 Pro（`seedance-1.5-pro`）を使う理由

Rust 12 stars

konjoai/vectro

⚡ Vectro: Lightning-fast embedding quantization. Hit 12M+ vec/s throughput (4.85× faster than FAISS C++) while drastically cutting memory footprint for vector databases and local AI research.

Overview: Vectro is a production-grade embedding compression library written in Rust, optimized for extreme compression with guaranteed quality. It delivers 12M+ vector throughput (4.85x faster than FAISS C++) while significantly reducing memory footprint for vector databases and local AI research. Key features: - **Ultra-High-Performance LLM Embedding Compressor** - **NEON 32-wide unroll** and Rayon block-coarsening for INT8 hot path - **Wave 1.3** profile flag for L2-normalized inputs - **Wave 2** fused single-pass kernel for efficient quantization - **Wave 3** SME2 and AVX-512-VNNI dispatch for improved performance - **Wave 3d** Apple Accelerate / AMX for d ≥ 256 multiply - **Wave 4** zero-copy and f16 entry point for improved performance - **Cross-platform packaging** via `cibuildwheel` - **Reproducibility harness** for consistent results Tech stack: Rust with Mojo SIMD acceleration Why it matters: Vectro is a highly optimized library for compressing large-scale vector embeddings, making it ideal for high-performance language models and AI research. Its performance benefits include up to 48x learned compression, native ANN search via HNSW index, and ultra-fast vector throughput, all while maintaining high cosine similarity accuracy. GitHub stars: 12

Python 12 stars

krylm/whisper-event-tuning

Final training script from HuggingFace Whisper Fine tuning event - to get best results on finetuned model.

### Overview: This project is a modified script for fine-tuning the Whisper model on the HuggingFace Fine Tuning Event 2022 dataset, aiming to optimize the model for better performance and faster training times. ### Key features: - Utilizes DeepSpeed for accessing larger batch sizes without gradient accumulation. - Concatenates audio and text samples to reduce the number of samples and improve model learning. - Implements SpecAugment for improved data augmentation in the Large v2 model. - Collects additional datasets for better training data, including Common Voice, Google FLEURS, and custom YouTube videos. ### Tech stack: Python ### Why it matters: This script aims to enhance the performance and efficiency of the Whisper model, making it more suitable for real-world applications and improving overall model accuracy.

Python 12 stars

kyopark2014/ocean-agent

It is a case study of an intelligent agent for Ocean.

### Overview: This project is a case study of an intelligent agent designed to analyze corporate information using an agentic workflow. The agent utilizes various PDFs to extract relevant information and summarize the content effectively. It employs metadata extraction and contextual embedding techniques to enhance the accuracy and completeness of the analysis. The project is implemented using Python and the LangChain library, with Amazon OpenSearch serving as the knowledge storage system. The architecture is designed to handle varying traffic and optimize costs, and it can be deployed using AWS CDK. ### Key Features: - Metadata extraction from PDFs to extract corporate names and creation dates. - Utilization of contextual embedding to summarize PDF chunks. - Integration of Amazon OpenSearch for efficient knowledge retrieval. - Multimodal handling to extract information from images and tables. - Hierarchical chunking to improve context and RAG search accuracy. - Use of Markdown format for report creation and conversion to HTML for external sharing. - Amazon serverless architecture for efficient handling of changing traffic and cost optimization. ### Tech Stack: - Python - LangChain - Amazon OpenSearch - AWS CDK ### Why It Matters: This project demonstrates the practical application of advanced AI techniques in corporate information analysis, showcasing how to efficiently extract and summarize large amounts of data from various sources. The use of Amazon OpenSearch and LangChain's OpenSearchVectorSearch provides a robust and scalable solution for knowledge management and retrieval, making it highly relevant in the modern business environment.

MDX 12 stars

lancedb/docs

Documentation for LanceDB, the Multimodal Lakehouse for AI

Overview: This project provides comprehensive documentation for LanceDB, a powerful multimodal lakehouse designed for AI applications. Key features: - Intuitive documentation - Clear explanations of features - Comprehensive examples - Easy-to-follow tutorials Tech stack: Markdown Why it matters: LanceDB is essential for developers looking to integrate multimodal data into their AI projects, offering a robust solution for data management and analysis.

Python 12 stars

le79pmv6mf/grok-imagine-image-prompts-es

Grok Imagine Image prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`grok-imagine-image`) and the cost per image.

### Overview: Grok Imagine Image Prompts: A Comprehensive Guide Explore a curated collection of 8 tested image prompt recipes, each meticulously crafted to ensure optimal rendering and cost-effectiveness. This repository offers a detailed look at the model ID (`grok-imagine-image`) and the cost per image, providing a practical guide for developers and creatives alike. ### Key Features: - **Model ID:** `grok-imagine-image` - **Recipes:** 8 (machine-readable in `data/prompts.json`, table in `PROMPTS.md`) - **Effective Price:** $0.016 per image - **Cost Estimate:** $0.128 for the entire set ### Tech Stack: Python ### Why It Matters: - **Cost Efficiency:** Ideal for developers looking to reduce costs while maintaining high quality. - **Model Compatibility:** Ensures seamless integration with OpenAI-compatible endpoints. - **Real-Time Pricing:** Provides up-to-date pricing information, enabling precise budgeting. ### Repository: le79pmv6mf/grok-imagine-image-prompts-es - **Stars:** 9 - **Description:** Grok Imagine Image prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`grok-imagine-image`) and the cost per image. - **Primary Language:** Python ### What Separates a Usable Recipe from a Wasted Call: - **Subject:** Name the material, not the adjective. - **Light:** Set the light explicitly. - **Framing:** Pin the framing with ratio and

Python 12 stars

lu60av/gemini-3-1-flash-image-preview-prompts-ru

Nano Banana 2 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-3.1-flash-image-preview`) and the cost per image.

### Overview: Nano Banana 2 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-3.1-flash-image-preview`), and the cost per image. ### Key features: - **Model ID:** `gemini-3.1-flash-image-preview` - **Aspect Ratios:** 1:1, 4:3, 8:5, 3:4, 16:9, 5:4, 1:2, 2:1 - **Cost:** $0.015 per image ### Tech stack: Python ### Why it matters: One key advantage of using Nano Banana 2 via APIMart is that it provides access to the same base URL and header, allowing for seamless integration with multiple models and APIs without the need for additional setup or billing. This makes it highly versatile and cost-effective for developers looking to enhance their image generation workflows.

Python 12 stars

marklubin/synix

Memory architectures that evolve with your agent

### Overview: Synix is a programmable memory architecture designed to evolve with your AI agent. It allows you to define and control the memory layers, prompts, and lifecycle of your agent's knowledge, ensuring accurate and contextually relevant information. ### Key Features: 1. **Pipeline Definition**: Define the memory layers, prompts, and lifecycle in Python. 2. **Sources**: Parse raw data from specified directories. 3. **Episodes**: Generate summaries for each source. 4. **Monthly Rollups**: Group episodes by month and synthesize them. 5. **Core Memory**: Compress all rollups into a single document. 6. **Search Index**: Index everything for full-text search. ### Tech Stack: Python ### Why It Matters: Synix provides a flexible and customizable memory architecture that allows you to evolve your agent's knowledge without the need for costly migrations or starting over. It ensures accurate and contextually relevant information, making it easier to manage and update your agent's memory as needed. ### GitHub Stars: 12 ### Repository: marklubin/synix ### Primary Language: Python ### Get Started: ```bash uvx synix init my-project cd my-project cp .env.example .env uvx synix build uvx synix release HEAD --to local uvx synix search "your query" --release local ``` ### What Just Happened: Synix processes your sources through a pipeline defined in `pipeline.py`. The pipeline includes: 1. Sources: Raw data parsed from `./sources/`. 2

Python 12 stars

martinwimpff/eeg-continual

Overview: This repository contains the code and data for a study on fine-tuning strategies for continual online EEG motor imagery decoding, focusing on insights from a large-scale longitudinal study. Key features: - Source training - Subject-specific fine-tuning - Causal evaluation with OTTA Tech stack: Python Why it matters: This repository provides a comprehensive framework for fine-tuning strategies in EEG motor imagery decoding, offering insights from a large-scale longitudinal study and facilitating reproducible results through checkpoints. Repository: martinwimpff/eeg-continual Primary language: Python GitHub stars: 12 GitHub description: No description provided. README excerpt: # Fine-Tuning Strategies for Continual Online EEG Motor Imagery Decoding: Insights from a Large-Scale Longitudinal Study This is the official repository for the paper [Fine-Tuning Strategies for Continual Online EEG Motor Imagery Decoding: Insights from a Large-Scale Longitudinal Study ](https://arxiv.org/pdf/2502.06828). ## Installation Run ``pip install .`` to install the ``eeg-continual`` package. ## Usage Reproducing the results is a 3-step process. To enhance reproducability, we have included [checkpoints](ckpts) for the first subject. ### 1. Cross-subject source training Run [source_training.py](src/eeg_continual/source_training.py) ``--config basenet.yaml`` to run the source training. *Note: You may want to change the [

Python 12 stars

mcqdlebg/gemini-2-5-flash-image-preview-api-gateway

Nano Banana API (`gemini-2.5-flash-image-preview`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Nano Banana API, also known as `gemini-2.5-flash-image-preview`, is a Nano Banana product offering a single OpenAI-compatible endpoint for image generation and editing at a cost of $0.0125 per image. It supports various resolutions and parameters, making it accessible and flexible for developers. Key features: - One key, entire catalog access via a single base URL and `Authorization` header. - Pay as you go pricing without the need for a monthly plan. - The price is reported in the response, so there's no guesswork. - Async by design, tasks can be submitted and monitored through the API. Tech stack: Python Why it matters: This API provides a straightforward and efficient way for developers to integrate image generation and editing capabilities into their applications, leveraging the Nano Banana product without the complexity of managing multiple endpoints or subscription plans. Repository: mcqdlebg/gemini-2-5-flash-image-preview-api-gateway GitHub stars: 8 GitHub description: Nano Banana API (`gemini-2.5-flash-image-preview`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

PowerShell 12 stars

mikecovlee/qwen3.8-27b-24gb-recipe

Run Qwen3.8-27B locally on any 24GB GPU (RTX 3090/4090/A5000) with llama.cpp: install, scripts, MTP tuning, benchmarks

Overview: This project provides a locally running setup for Qwen3.8-27B, a 24GB GPU version, using llama.cpp. It includes two recipes, one with vision support and the other without, catering to different VRAM budgets and vision needs. Key features: - Two recipes available, each with its own unique features. - Full 256K native context, no rope scaling, and q4_0 KV cache. - MTP speculative decoding for faster generation. - Optional GPU vision support. - OpenAI-compatible API for compatibility with any OpenAI-compatible client. - Resumable, size-verified downloads for easy installation. - One-command installs for resumable and size-verified downloads. Tech stack: PowerShell and llama.cpp. Why it matters: This setup allows developers to run Qwen3.8-27B locally on any 24GB GPU, including the RTX 3090, 4090, and A5000, with the option to use vision support for better performance. Repository: `mikecovlee/qwen3.8-27b-24gb-recipe` Primary language: PowerShell GitHub stars: 12 GitHub description: Run Qwen3.8-27B locally on any 24GB GPU (RTX 3090/4090/A5000) with llama.cpp: install, scripts, MTP tuning, benchmarks README excerpt: A ready-to-run setup for serving Qwen

Unknown 12 stars

mlnjsh/context-engineering

🧠 The Complete Guide to Context Engineering — Research Papers, Blogs, YouTube Videos, Tools & Best Practices for LLM Context Window Optimization

Overview: Context Engineering is the process of designing, managing, and optimizing the information provided to Large Language Models (LLMs) within their context window to enhance the quality, accuracy, and relevance of their outputs. Key features: - Context Window Sizes: Determining the optimal size of the context window to balance between relevance and efficiency. - Research Papers: A curated list of research papers on Context Engineering. - YouTube Videos & Talks: A collection of videos and talks related to Context Engineering. - Blog Posts & Articles: Insights and discussions on Context Engineering from various sources. - Tools & Frameworks: Tools and frameworks used in Context Engineering. - Courses & Tutorials: Educational resources for learning Context Engineering. - Techniques & Patterns: Strategies and best practices for Context Engineering. - Roadmap: A roadmap for the future development of Context Engineering. - Contributing: Guidelines for contributing to the project. - Citation: Instructions for citing the project. Tech stack: The project uses Markdown, GitHub, and various tools for documentation and collaboration. Why it matters: Context Engineering is crucial for improving the performance and quality of Large Language Models by optimizing the information provided to them within their context window. This ensures that the models can generate more accurate and relevant responses, leading to better user experiences and improved applications.

Python 12 stars

muaaz-ur-habibi/fthedocs

A documentation querying engine, useful for scanning the docs in a question-answer like manner

Overview: FTheDocs is a documentation querying engine designed to answer questions about software documentation efficiently. It uses a vector database approach to find the closest matching information from a set of documents. Key features: - Vector database querying - Question-Answer style querying system - Customizable settings - Free and open-source software Tech stack: Python Why it matters: FTheDocs simplifies the process of finding answers in software documentation, making it easier and faster to understand complex systems. Repository: muaaz-ur-habibi/fthedocs Primary language: Python GitHub stars: 12 GitHub description: A documentation querying engine, useful for scanning the docs in a question-answer like manner

Python 12 stars

muqadasejaz/PDF-QA-RAG-System

A PDF Question-Answering App built with RAG (Retrieval-Augmented Generation), allowing users to upload PDFs and ask context-based questions. Powered by Streamlit, LangChain, Ollama, and Chroma for efficient and accurate answers.

Overview: This project is a PDF Question-Answering App built with RAG (Retrieval-Augmented Generation), allowing users to upload PDFs and ask context-based questions. Powered by Streamlit, LangChain, Ollama, and Chroma for efficient and accurate answers. Key features: - Upload PDF and extract text automatically - Context-aware Q/A using RAG - Efficient Vector Search with Chroma/FAISS backend - LLM-powered answers using Ollama + Llama 3.1 - Interactive GUI built with Streamlit - Handles temporary files securely during processing Tech stack: Python, Streamlit, LangChain, Ollama, Chroma/FAISS Why it matters: This project provides a user-friendly solution for extracting and answering questions from PDFs, leveraging advanced AI techniques to ensure accurate and context-aware responses.

JavaScript 12 stars

neomjs/neo-agent-brain

Point a cross-model AI engineering team at your own codebases. The Agent OS keeps named AI maintainers' identity, memory and review across labs alive between sessions — the Brain Neo.mjs runs on itself. Multi-tenant: onboarding a codebase is a config entry, not a fork.

Overview: The Brain of the Neo.mjs organism, the Agent OS, is the central repository that provides persistent identity, shared memory, and review discipline for the AI engineering team. Key features: - Persistent identity for AI maintainers - Durable institutional memory - Peers from different model families - Right to refuse and challenge - Shared standard and human merge gate Tech stack: JavaScript Why it matters: The Brain ensures accountability, reliability, and accountability across multiple AI agents, creating a stable and transparent engineering environment.

Python 12 stars

p61qrgpg/gemini-3-1-flash-image-preview-prompts-id

Nano Banana 2 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-3.1-flash-image-preview`) and the cost per image.

### Overview: This repository contains Nano Banana 2 prompts designed for image and video rendering. The prompts are tested on the `gemini-3.1-flash-image-preview` model and provide detailed examples and cost estimates. ### Key Features: - **Model ID:** `gemini-3.1-flash-image-preview` - **Prompts:** 8 recipes with examples and cost estimates - **Aspect Ratios:** 1:1, 4:3, 8:5, etc. - **Cost:** $0.015 per image (estimated cost for the set: $0.12) ### Tech Stack: - Python ### Why It Matters: This repository offers a practical guide for developers looking to create high-quality images and videos using Nano Banana 2, providing clear instructions and cost estimates.

Shell 12 stars

parth-thakre/exl3-rdna4-recipe

Qwen3.8-27B (EXL3) on an AMD RX 9070 XT: ~150 tok/s for code, 96k context, one-line install (container, any Linux distro)

Overview: This project is a Qwen3.8-27B model optimized for an AMD RX 9070 XT, running on Linux. It includes ExLlamaV3, TabbyAPI, and RDNA4 patches, providing high performance and context for coding and chat tasks. Key features: - High throughput for code and short context, ~150 tokens per second. - Accurate speculative decoding with a 96k context by default. - Supports 128k context for larger tasks. - Multiple requests can be handled simultaneously. - One-line installation for easy setup. Tech stack: Shell-based installation and run script. Why it matters: This model offers exceptional performance for coding and chat tasks, making it suitable for developers and AI enthusiasts looking for high-speed, accurate, and context-rich solutions. Repository: parth-thakre/exl3-rdna4-recipe Primary language: Shell GitHub stars: 10 GitHub description: Qwen3.8-27B on an AMD RX 9070 XT: ~150 tok/s for code, 96k context, one-line install (container, any Linux distro)

TypeScript 12 stars

patelvivekdev/turso-vector-search

A demo AI chatbot with Vercel AI SDK, Turso as vector DB and google gemini model.

Overview: This project is a demo AI chatbot using Turso Vector Search, Vercel AI SDK, and Google Gemini model for vector database integration. Key features: - Turso Vector Search API integration - Drizzle ORM for database operations - Google Gemini model for AI-based responses - Tailored UI for enhanced interaction Tech stack: TypeScript Why it matters: This project showcases the integration of AI and vector search technologies, demonstrating how they can be used to create sophisticated chatbots and AI-driven applications. Repository: patelvivekdev/turso-vector-search Primary language: TypeScript GitHub stars: 12 GitHub description: A demo AI chatbot with Vercel AI SDK, Turso as vector DB and google gemini model.

Swift 12 stars

ph1lb4/imagegen-mac

Local AI image generation for Mac. Run Qwen-Image 2.1 on Apple Silicon, offline and private, with a built-in MCP server for Claude Code and AI agents.

### Overview: ImageGen is a local AI image generation tool for macOS that runs on Apple Silicon, providing offline and private image generation capabilities. It uses Qwen-Image 2.1 and a built-in MCP server for Claude Code and other AI agents, making it suitable for developers and creatives. ### Key Features: - **Fully on-device**: Prompts and images never leave your Mac. - **Text to image**: Strong typography with readable text inside images. - **Image editing**: Change images with instructions or combine up to 10 reference images. - **Transparent PNGs**: Native RGBA output for stickers, icons, and assets. - **MCP server for AI agents**: Supports Claude Code and other MCP clients for image generation and editing. - **One-click model download**: From Hugging Face, resumable within the app. - **Memory guards**: Prevents big jobs from freezing your Mac. - **Native SwiftUI app**: Menu bar icon, keeps the engine running even when the window is closed. ### Tech Stack: SwiftUI ### Why It Matters: - **State-of-the-art model**: Fits in about 33 GB of memory, making it ideal for Mac users. - **Local image generation**: No need for cloud services or API keys, eliminating monthly bills. - **Native performance**: Runs on your Mac's GPU (Metal) for fast and efficient image generation. - **Built-in MCP server**: Supports Claude Code and other AI agents, allowing for seamless image creation and editing. ### Gallery: - **Neon Sign**: A neon sign reading "