Project directory

Projects

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

Python 17 stars

atpfffifl97hm/flux-kontext-api-pt

FLUX Kontext API (`flux-kontext-pro`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: FLUX Kontext API is an OpenAI-compatible API endpoint that provides model IDs, pricing, and parameters for generating images, videos, and language content. It allows developers to access a catalog of over 300 models using a single API endpoint. Key features: - Model ID and pricing information - Access to over 300 models via a single API endpoint - Pricing per image, with flexible resolution options - Assynchronous nature for efficient use - Real-time pricing updates Tech stack: Python Why it matters: This API is crucial for developers looking to integrate OpenAI-compatible image generation into their projects without the need for a monthly subscription, making it cost-effective and scalable. Repository: atpfffifl97hm/flux-kontext-api-pt Primary language: Python GitHub stars: 17 GitHub description: FLUX Kontext API (`flux-kontext-pro`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Unknown 17 stars

atryx/ai-engineer-roadmap-2026

AI engineer roadmap for 2026 — LLMs, RAG, agents, fine-tuning, vector databases, prompt engineering, MLOps, and building production AI apps

Overview: This AI engineer roadmap is designed for professionals aiming to become a production AI engineer in 2026. It covers key areas such as LLM fundamentals, prompt engineering, RAG, AI agents, fine-tuning, evaluation, and deploying AI-powered applications. Key features: - LLM fundamentals (LLM, tokens, context windows, temperature, embeddings) - Prompt engineering - RAG (Retrieval-Augmented Generation) - AI agents - Fine-tuning and training - Evaluation and testing - Production AI systems - Multi-modal AI - Advanced patterns Tech stack: Python, Jupyter notebooks, virtual environments, JSON processing, SQL, pandas, linear algebra, probability, statistics, APIs, and machine learning models. Why it matters: This roadmap provides a structured approach to building, deploying, and operating AI-powered applications using modern AI tools and techniques. Repository: atryx/ai-engineer-roadmap-2026 GitHub stars: 17 GitHub description: AI engineer roadmap for 2026 — LLMs, RAG, agents, fine-tuning, vector databases, prompt engineering, MLOps, and building production AI apps. README excerpt: This roadmap is for AI Engineers — you build products on top of models, not train them from scratch. It covers foundational skills, LLM fundamentals, prompt engineering, RAG, AI agents, fine-tuning, evaluation, and deployment. Project Ideas by Level: The roadmap includes project ideas categorized by level, providing a clear path for learning and development.

Python 17 stars

aws-samples/rag-with-amazon-postgresql-using-pgvector-and-sagemaker

Question Answering application with Large Language Models (LLMs) and Amazon Postgresql using pgvector

### Overview: This project is a Question Answering application that leverages Large Language Models (LLMs) and Amazon Postgresql using pgvector. It utilizes the RAG (Retrieval Augmented Generation) approach to retrieve the most relevant information from the enterprise knowledge base or content, enhancing the accuracy of the LLM responses. ### Key Features: - **PGVector Integration**: Amazon Aurora Postgresql with pgvector for knowledge base retrieval. - **RAG Architecture**: Combining LLM with pre-generated embeddings for improved accuracy. - **Deployment**: SageMaker Studio, SageMaker Endpoints, Amazon Aurora Postgresql, and AWS Secrets Manager. ### Tech Stack: - Python ### Why It Matters: Amazon Aurora Postgresql with pgvector is used for knowledge base retrieval, optimizing generative AI applications, and improving the performance of generative AI workloads on Amazon Aurora.

Python 17 stars

batuhanzorbeyzengin/rebasis

Measure whether an embedding upgrade is worth it, bridge it without reindexing when it is, and migrate safely when you are ready.

### Overview: rebasis is a tool designed to help developers measure the value of embedding upgrades without the need for a full reindex. It bridges the upgrade seamlessly, ensuring that the index remains unchanged while still providing meaningful results. ### Key Features: 1. **Decision on Upgrade Worthiness:** rebasis decides whether an upgrade is worth performing based on the retention rate of the new model's embeddings compared to a full reindex. 2. **Query-Based Alignment:** It aligns the new model's query vectors with the existing index space, ensuring that the index remains unchanged. 3. **Bridge and Migrate Safely:** rebasis allows developers to bridge the upgrade without reindexing, providing a safe migration process. ### Tech Stack: rebasis is built with Python, utilizing the PyTorch framework for its core functionality. ### Why It Matters: rebasis provides a practical solution for developers who need to upgrade their embeddings without disrupting their existing infrastructure. By deciding whether an upgrade is worth it based on query results, rebasis helps developers make informed decisions, ensuring that they only upgrade when necessary.

Jupyter Notebook 17 stars

bilgeyucel/multimodal-agent-workshop

🖼️ Workshop: Build a multimodal AI agent with Haystack & GPT-4o — featuring image understanding, document retrieval, conversational memory, and human-in-the-loop safety controls

Overview: Dive into the world of multimodal AI with this workshop, where you'll build a sophisticated agent that integrates image understanding, document retrieval, conversational memory, and human-in-the-loop safety controls using Haystack and GPT-4o. Key features: - Multimodal indexing pipeline for PDFs and images - Vision-enabled GPT-4o agent - RAG tool for company policy retrieval - Conversational memory for context-aware interactions - Human-in-the-loop controls for sensitive actions Tech stack: Python 3.10+ with Haystack and GPT-4o integration. Why it matters: This workshop provides hands-on experience in building a multimodal AI agent, showcasing cutting-edge technologies like CLIP embeddings, GPT-4o, and Haystack, making it ideal for developers looking to enhance their AI capabilities in various applications.

Python 17 stars

cahlen/conversation-dataset-generator

Generate multi-speaker conversational datasets for LLM fine-tuning. N-persona support, topic variation, conversation continuation, creative briefs with web search, character pools, Docker support. ShareGPT format with configurable role mapping.

Overview: This project generates synthetic conversational datasets in the ShareGPT format for fine-tuning large language models (LLMs). It supports multi-persona, topic variation, conversation continuation, and creative briefs with web search context. The datasets are optimized for use with LLMs and can be configured with various parameters for customization. Key features: - Multi-persona support - Topic variation - Conversation continuation - Creative briefs with web search context - Character pools - Docker support Tech stack: Python Why it matters: This project is essential for developers looking to fine-tune LLMs for natural language understanding and generation tasks. The ability to generate diverse and realistic conversations tailored to specific use cases is crucial for improving model performance and application effectiveness. Repository: cahlen/conversation-dataset-generator Primary language: Python GitHub stars: 17 GitHub description: Generate multi-speaker conversational datasets for LLM fine-tuning. N-persona support, topic variation, conversation continuation, creative briefs with web search, character pools, Docker support. ShareGPT format with configurable role mapping.

Python 17 stars

castacks/tartanmatch

TartanMatch: universal dense correspondence across RGB, depth, thermal, LiDAR, and event modalities

### Overview: TartanMatch is a universal dense correspondence model that predicts optical flow and covisibility between any two views captured in RGB, depth, thermal, LiDAR, or event modalities. ### Key features: - Predicts optical flow and covisibility for every pixel in the source and target images. - Supports 25 modality pairs: RGB, depth, thermal, LiDAR, and event. - Uses a shared DINOv2 encoder and multi-view global-attention transformer. - Maps non-RGB modalities into the encoder's image space. - Supports arbitrary source and target sizes and resolutions. - Predicts flow and covisibility in the original pixel units of the source and target images. ### Tech stack: The project is implemented in Python using the following libraries: - PyTorch - NumPy - Safetensors - Hugging Face Hub ### Why it matters: TartanMatch enables the dense correspondence between any two views captured in different modalities, making it a versatile tool for applications such as scene reconstruction, 3D modeling, and scene analysis.

Python 17 stars

dkylewillis/vera

Overview: VERA is a vector-embedded retrieval archive that converts documents into searchable `.vera` files, making them easily accessible and searchable across different formats and platforms. Key features: - Converts documents into `.vera` files. - Supports multiple formats including PDF, DOCX, PPTX, and Excel. - Provides keyword and semantic search modes. - Automatically updates and indexes archives when files change. - Supports local and offline usage with a local model runtime. Tech stack: Python. Why it matters: VERA simplifies document management by converting and indexing documents into searchable archives, making them easily accessible and searchable across different formats and platforms.

TypeScript 17 stars

edisonneza/image-to-text

PWA - Convert Image to Text - A small multi language project built to use the frontend libraries and create an offline app that can be installed in any device.

Overview: This project is a Progressive Web App (PWA) that converts images to text using JavaScript and the Tesseract OCR library. It supports multiple languages and can be installed on any device. Key features: - Paste image from Clipboard - Add new option to convert images in backend - Optimize speed using more workers - Drag and Drop feature Tech stack: TypeScript Why it matters: This project demonstrates the use of frontend libraries and creates an offline app that can be installed on any device, making it useful for various applications requiring image-to-text conversion. Repository: edisonneza/image-to-text Primary language: TypeScript GitHub stars: 17

Jupyter Notebook 17 stars

enoche/DGVAE

Disentangled Graph Variational Auto-Encoder for Multimodal Recommendation with Interpretability, IEEE TMM

### Overview: The Disentangled Graph Variational Auto-Encoder (DGVAE) is a novel approach designed to enhance the interpretability of multimodal recommendation systems by disentangling the representations of users and items. This method leverages state-of-the-art multimodal pre-training techniques to convert images into textual content, constructs a frozen item-item graph, and encodes interactions into disentangled representations. The model then regularizes these representations through mutual information maximization, facilitating interpretability through text-based user interactions. ### Key Features: - **Disentangled Representation Learning**: The model disentangles the representations of users and items to enhance interpretability. - **Multimodal Information Utilization**: It utilizes state-of-the-art multimodal pre-training technologies to convert images into textual content. - **Frozen Item-Item Graph**: Constructs a frozen item-item graph to represent interactions between items. - **Regularized Disentangled Representations**: Aligns disentangled representations derived from interactions with those learned from textual content, facilitating interpretability through text-based user interactions. ### Tech Stack: - **Programming Languages**: Jupyter Notebook - **GitHub Stars**: 17 ### Why It Matters: The performance of `DGVAE` significantly surpasses that of state-of-the-art baselines, achieving a margin of 10.02% in accuracy metrics on real-world datasets. This method offers a practical solution for enhancing the interpretability of multimodal recommendation systems, making it particularly valuable in applications where transparency and understanding of user interactions are crucial.

Python 17 stars

firstbatchxyz/firstbatch-sdk

Python SDK for FirstBatch: Real-time personalization using vectorDBs

Overview: The FirstBatch SDK is a Python interface designed to integrate vector databases and power real-time personalized AI experiences in applications. Key features: - Seamless management of user sessions with persistent IDs or temporary sessions - Real-time updates of user embeddings through signal actions - Personalized data fetching tailored to each user's embeddings - Support for multiple vector database integrations: Pinecone, Weaviate, etc. - Built-in algorithms for common personalization use cases - Easy configuration with Python classes and environment variables Tech stack: Python Why it matters: The FirstBatch SDK simplifies the integration of vector databases and personalization algorithms, enabling developers to create highly personalized user experiences without the complexity of low-level vector database management.

Ruby 17 stars

flanker/chroma-db-ui

A simple Ruby UI for Chroma database.

Overview: A simple Ruby UI for Chroma database. Key features: - User-friendly interface for managing Chroma database - Integration with Chroma Ruby Client for efficient database operations Tech stack: Ruby on Rails Why it matters: This tool simplifies the management of Chroma database, making it accessible and easy to use for developers and administrators alike.

Python 17 stars

instavar/cosyvoice3-lora-finetuning

CosyVoice3 LoRA fine-tuning companion repo — PEFT integration, 9 known pitfalls from IMDA NSC production runs, corrected from failed full-SFT attempt

### Overview: CosyVoice3 LoRA Fine-Tuning Companion Repo This repository provides comprehensive tools and scripts for fine-tuning the CosyVoice3 model using LoRA (Low-Rank Adaptation). It includes PEFT integration, selective layer freezing, LoRA-aware checkpoint management, and overfitting detection. The companion repo is designed for single-speaker voice cloning on a consumer-grade GPU, leveraging the Qwen2-based LLM backbone. ### Key Features: - **LoRA Training with PEFT**: Integrates LoRA with the Qwen2-based LLM backbone. - **Selective Layer Freezing**: Configurable unfreezing of specific layers. - **LoRA-Aware Checkpoint Save/Load**: Saves only the LoRA-aware adapters. - **Overfitting Detection**: Validates the training loop for overfitting. ### Tech Stack: - **Python**: The primary programming language used for the project. - **DeepSpeed**: A library for training large models efficiently. ### Why It Matters: CosyVoice3 is an advanced voice cloning model that supports fine-tuning using LoRA. This repo provides all necessary components for such fine-tuning, including PEFT integration, selective layer freezing, and LoRA-aware checkpoint management. It also includes evaluation and checkpoint management scripts, making it a valuable resource for developers working with CosyVoice3. ### Repository Details: - **Repository**: instavar/cosyvoice3-lora-finetuning - **Stars**: 17 - **GitHub Description**: Companion repo for single-speaker voice cloning on a

Python 17 stars

irermc20d41/wan2-7-prompts

Wan 2.7 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`wan2.7`) and the cost per image.

### Overview: Wan 2.7 Prompts: A Comprehensive Guide to Image Prompt Recipes ### Key Features: - **Model ID:** `wan2.7` - **Cost Per Image:** $0.0664 per second - **Aspect Ratios:** 16:9, 1:1, 3:4, etc. - **Example Renders:** Product Ad, Cinematic, Food Ad, etc. ### Tech Stack: - Python ### Why It Matters: - **Efficient Pricing:** $0.0664 per second, no subscription required. - **Pay-As-You-Go Model:** No monthly plan, no free quota. - **Immediate Billing:** Every call reports the cost, ensuring accurate spending tracking. ### Repository: - **Repository:** irermc20d41/wan2-7-prompts - **GitHub Stars:** 17 ### Primary Language: - **Primary Language:** Python ### GitHub Description: Wan 2.7 Prompts: Tested Image Prompt Recipes with Example Renders, Aspect Ratios, the Model ID (`wan2.7`), and the Cost Per Image. ### Gallery: Each row represents a different prompt, including example renders, category, duration, and cost. - **Product Ad:** 5 seconds, $0.0664, "Slow 180° orbit around a matte ceramic mug on a dark stone plinth, steam rising, soft key light from camera left, shallow depth of field, product commercial." - **

Python 17 stars

iu6iw6cxi4z2in/qwen-image-2-0-prompts-cn

Qwen-Image 2.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`qwen-image-2.0`) and the cost per image.

### Overview: Qwen-Image 2.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`qwen-image-2.0`), and the cost per image. ### Key features: - **Model ID:** `qwen-image-2.0` - **Aspect Ratios:** 1:1, 4:3, 16:9 - **Cost Per Image:** $0.02 ### Tech Stack: Python ### Why it Matters: - **Cost Efficiency:** $0.02 per image, no need for monthly subscriptions. - **OpenAI-Compatible Endpoint:** `https://api.apimart.ai/v1`, supports various models. - **Real-Time Pricing:** Returns actual cost upon completion, refunds failed tasks. - **Asynchronous Design:** Batch processing and retry handling without custom integrations. ### Repository: - Repository: iu6iw6cxi4z2in/qwen-image-2-0-prompts-cn - GitHub Stars: 8 - GitHub Description: Qwen-Image 2.0 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`qwen-image-2.0`) and the cost per image.

Python 17 stars

j-az78q90oa/seedance-2-0-fast-api-fr

Seedance 2.0 Fast API (`doubao-seedance-2.0-fast`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Seedance 2.0 Fast API is an OpenAI-compatible endpoint for generating text-to-video and image-to-video content, billed per generated second. It offers a single endpoint compatible with OpenAI, without the need for a monthly subscription. Key features: - Single endpoint compatible with OpenAI - Text-to-video and image-to-video functionalities - Billing per generated second - Supports multiple resolutions and framing options - Asynchronous processing Tech stack: Python Why it matters: Seedance 2.0 Fast API simplifies the process of integrating Seedance 2.0 with OpenAI-compatible models, providing a straightforward and efficient solution for developers without the need for a monthly subscription.

Python 17 stars

j2pnwx3z8juf/nano-banana-pro-api-cn

香蕉 — Nano Banana Pro API (`gemini-3-pro-image-preview`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: Nano Banana Pro API is an OpenAI-compatible endpoint offering image generation and editing services at a cost of $0.03 per image, with a 4K resolution option available. It supports a wide range of models and is priced per image, with a minimum charge of $1 for the first use. Key features: - Supports 4K image generation and editing - Pricing model: $0.03 per image, $0.04 for 4K - Model id: `gemini-3-pro-image-preview` - Model page: [model page](https://go.apimart.ai/k-8d5e43) - Billing model: per image, with a minimum of $1 for the first use - Model id and routes: `gemini-3-pro-image-preview` - Pricing: $0.03 per image, $0.04 for 4K - Request parameters: `model`, `prompt`, `size`, `resolution`, `n`, `image_urls`, `nsfw_check` - Model page: [model page](https://go.apimart.ai/k-8d5e43) - Model id and routes: `gemini-3-pro-image-preview` - Pricing: $0.03 per image, $0.04 for 4K - Request parameters: `model`, `prompt`, `size`, `resolution`, `n`, `image_urls`, `nsfw_check` - Model page: [model page](https://go.apimart

Swift 17 stars

joeynyc/honeycomb-lab

Honeycomb Lab — hex map + OpenAI gateway control plane for a home AI fleet

Overview: Honeycomb Lab is a control plane and OpenAI-compatible gateway designed to manage and monitor your home AI fleet, providing a unified API endpoint for all models and real-time traffic visualization. Key features: - A single API endpoint for all OpenAI-compatible clients - A live map that tracks the health and performance of your GPUs - Fleet control capabilities from the map, including PING, SERVE/STOP, and DOCTOR - Works seamlessly on macOS, iPad, and phone, with a native app and a self-contained web dashboard Tech stack: Swift, Python (for the gateway), and a hex map visualization Why it matters: Honeycomb Lab simplifies the management and monitoring of your AI fleet, making it easier to route models, track performance, and control your fleet from a single interface, enhancing efficiency and reliability.

Python 17 stars

jpblmt-we/flux2-prompts-de

FLUX 2 Pro prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`flux-2-pro`) and the cost per image.

Overview: This repository contains tested image prompt recipes for the FLUX 2 Pro model, including example renders, aspect ratios, the model ID (`flux-2-pro`), and the cost per image. Key features: - 8 tested recipes - Example renders, ratios, and cost per image - Model: `flux-2-pro` - Prompt structure cheat-sheet Tech stack: Python Why it matters: Provides a comprehensive guide for developers using the FLUX 2 Pro model, ensuring accurate and efficient image generation.

Python 17 stars

js-lee-AI/MADARA

Retrieval-augmented generation: training-free document-assessment routing for cost-efficient multi-agent RAG.

### Overview: MADARA is a training-free pipeline designed to diagnose and route model-document assessments for cost-efficient multi-agent retrieval-augmented generation (RAG). It identifies the most cost-effective treatment for each model-task pair, ensuring accurate and efficient document assessment. ### Key Features: - **RSC (Reasoning-Score Coupling)**: A label-free probe that perturbs reasoning in three steps to check for monotonic degradation of scores. - **NF (No-Filter)**: Measures the model's raw context-handling capacity. - **Routing**: - **PDE (Per-Document Extraction)**: Answers from each top-k document in isolation and votes. - **SDA (Score Distribution Alignment)**: Converts scores to percentile ranks before aggregation. - **CoT De-Polarization**: Forces step-by-step reasoning before scoring. - **ATF (Adaptive Threshold Filtering)**: Keeps threshold filtering for models with extreme polarization. ### Tech Stack: MADARA is implemented using Python and Mistral-7B. ### Why It Matters: MADARA simplifies the cost of multi-agent RAG by identifying the most efficient document assessment treatment for each model-task pair, ensuring accurate and efficient document assessment.