Projects

Overview: SochDB is an embedded, AI-native database designed specifically for AI agents and memory management, providing a single engine and file for structured data, embeddings, and agent memory, eliminating the need for multiple databases and integration layers. Key features: - SQL-92-compatible with `JOIN`s, aggregates (`GROUP BY` / `SUM` / `AVG` / `HAVING`), and MySQL/PostgreSQL/SQLite dialect normalization - HNSW vectors fused with BM25 via Reciprocal Rank Fusion for hybrid search - Bi-temporal knowledge graph with point-in-time recall - Context Query Builder for multi-source fusion under a token budget, with TOON dense output - Full ACID transactions with MVCC + WAL + Serializable Snapshot Isolation Tech stack: Rust Why it matters: SochDB simplifies the complexity of integrating multiple databases and retrieval layers, providing a unified and efficient solution for AI agents and memory management.

Overview: Sora JSON Prompt Crafter is a web interface designed for crafting Sora's generative models' prompts, offering a user-friendly interface with sliders, dropdowns, and privacy-first storage to generate and customize JSON snippets directly for Sora. Key features: - Live JSON generation with diff highlighting - Edit prompts and negative prompts - Privacy-first local storage - Dark mode - 100% locally generated JSON - JSON composition based on option sections - JSON generation lifecycle functions - Persisted current generated JSON - History panel for up to 100 copied JSON prompts - Import and export functions - Dedicated settings panel for customization - Presets for quality, ambient, and settings - Pop culture presets - Specific framing, dimension, and base presets - Video camera and motion presets - Advanced specialized prompting options - Artifact and defect correction presets - No-fuss tracking toggle - Shared links for referral tracking - Internationalization support - Works offline with service worker caching - Updates service worker on load and Settings panel - Inject custom CSS and JavaScript from Settings panel Tech stack: TypeScript Why it matters: Sora JSON Prompt Crafter simplifies the process of crafting prompts for Sora's generative models, making it accessible and efficient for both developers and prompt engineers. The web interface and privacy-first approach ensure that user data is kept secure and only local, enhancing user experience and satisfaction.

Overview: Z-Image Turbo API is a fast text-to-image service on APIMart, priced per image, offering a single key access to over 300 models and flexible pricing options. Key features: - Supports multiple models and resolutions - Pricing is per image, with a minimum of $1 - Charges are reported in the response, and tasks can be refunded if failed - Supports async processing with task IDs Tech stack: Python Why it matters: It's a cost-effective and flexible way to access a wide range of image models without the need for a monthly subscription. Repository: t9hid1p6ue2a/z-image-turbo-api-ja Primary language: Python GitHub stars: 47 GitHub description: Z-Image Turbo API (`z-image-turbo`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.

Overview: TripletCLIP is an official PyTorch implementation of the NeurIPS 2024 paper "Improving Compositional Reasoning of CLIP via Synthetic Vision-Language Negatives". This repository provides access to the dataset, pretrained checkpoints, inference, and training code for the project. Key features: - TripletData dataset - Pretrained and finetuned checkpoints - TripletCLIP adaptation on OpenCLIP - Data generation scripts - Full TripletData release - Original TripletCLIP training scripts for reproducibility Tech stack: Python Why it matters: This implementation provides a comprehensive resource for researchers and developers interested in improving compositional reasoning in CLIP models, making it easier to reproduce results and build upon the work of the NeurIPS 2024 paper.

Overview: FedAudio is a benchmark library for evaluating federated learning methods on audio tasks, presented at ICASSP 2023. It includes FL feature managers, pre-processing managers, and data splitters, providing a realistic benchmark for algorithmic research. Key features: - FL feature managers to simulate real-world federated learning challenges - Pre-processing manager supporting various audio features and deep audio representations - Data splitter for different types of data partitioning Tech stack: Python Why it matters: FedAudio offers a comprehensive benchmark for federated learning on audio tasks, addressing the need for realistic benchmarks in the field.

Overview: RAG-ARC is a modular, high-performance Retrieval-Augmented Generation framework designed for efficient, scalable architectures that support multi-path retrieval, graph extraction, and fusion ranking. Key features: - Multi-Format Document Parsing - Text Chunking & Vectorization - Multi-Path Retrieval - Graph Structure Extraction - GraphRAG - Re-ranking (Rerank) - Modular Design Tech stack: Python Why it matters: RAG-ARC addresses key challenges in traditional RAG systems, ensuring high recall and semantic consistency for knowledge-intensive tasks, making it suitable for academic research, personal knowledge bases, and enterprise-level knowledge management systems. Repository: DataArcTech/RAG-ARC GitHub stars: 48

Overview: This is a Python module designed to inspect and convert coding-agent session files (`.jsonl`) — Claude Code, Codex, and Pi — into the messages format expected by Hugging Face Transformers. Key features: - Inspects and converts Claude Code, Codex, and Pi session files into messages format. - Handles stage 1 of the conversion pipeline, which parses the JSONL files into a simple `[{"role": ..., "content": ...}]` list. - Supports three different tools: Claude Code, Codex, and Pi. - Provides a unified `Converter` class for easy access to all three tools. - Supports installation via `pip install claude-converter`. - Offers a quick start guide for each tool, including example usage. Tech stack: Python. Why it matters: This module simplifies the process of working with session files from Claude Code, Codex, and Pi, making it easier to integrate these tools into workflows that require the Hugging Face Transformers format. Repository: FredyRivera-dev/claude_converter GitHub stars: 48 GitHub description: A zero-dependency Python module for inspecting and converting coding-agent session files (`.jsonl`) — Claude Code, Codex, and Pi — into the `messages` format expected by Hugging Face Transformers. README excerpt: A zero-dependency Python module for inspecting and converting coding-agent session files (`.jsonl`) — Claude Code, Codex, and Pi — into the `messages` format expected by Hugging Face Transformers.

### Overview: This repository contains an awesome list of Vision Language (VL) Prompt, Finetune, and Adapter methods and models. The papers are organized chronologically, providing a comprehensive overview of the development of this field. The list includes methods for efficient adaptation of pre-trained vision-language models to various tasks and datasets. ### Key Features: - Dual Modality Prompt Tuning (DPT) - Scaling & Shifting Your Features: A New Baseline for Efficient Model Tuning (SSF) - Debiasing Vision-Language Models via Biased Prompts (Debias_VL) - PLOT: Prompt Learning with Optimal Transport for Vision-Language Models (PLOT) ### Tech Stack: - CLIP (Pre-trained Model) - CoOp (Pre-trained Model) - Optimal Transport (OT) Theory ### Why it Matters: This repository serves as a valuable resource for researchers and practitioners interested in the efficient adaptation of pre-trained vision-language models. It provides a systematic overview of the latest methods and models, facilitating a deeper understanding of the field and enabling the development of more effective and efficient solutions.

### Overview: 该项目利用ROS2 Humble框架结合Gazebo环境，实现通过语音命令控制轮式机器人的视觉目标检测、SLAM导航和抓取物体的功能。项目使用了YOLOv8进行目标检测和大模型进行语音指令识别，通过多任务分支实现复杂任务的执行。 ### Key features: - **语音指令识别**：通过大模型将语音命令转换为指令，控制机器人完成目标检测、建图导航和抓取物体。 - **多任务分支**：在基础任务分支的基础上，增加了角度校准和场景航点，支持多地点抓取任务。 - **机械臂控制**：通过moveit动作组规划控制机械臂进行抓取操作。 ### Tech stack: Python ### Why it matters: 该项目展示了如何利用ROS2和Gazebo框架实现复杂的机器人控制任务，特别是通过语音命令和视觉识别技术，实现了机器人在不同场景下的自主导航和目标抓取，具有广泛的应用前景和研究价值。

Overview: Monan is a native AI Agent framework for Bun, designed to simplify the creation of secure and efficient local-first agents. It leverages Bun's native architecture for local inference, SQLite vector storage, and workflow orchestration. Key features: - Local inference with bare-metal performance - Embedded vector search for instant API deployment - Secure agent creation with minimal latency - Built with TypeScript and supports Bun Tech stack: TypeScript Why it matters: Monan simplifies the process of creating AI agents, eliminating latency and complexity, making it ideal for developers building local-first applications. Its integration with Bun's native architecture ensures optimal performance and ease of use. Repository: Monan-AI/monan-sdk GitHub stars: 48 GitHub description: The native AI Agent framework for Bun. Build local-first, bare-metal agents with embedded vector search and instant API deployment. 🚀

Overview: Void is a voice-first desktop companion for Claude Code, allowing users to interact with Claude Code through their computer's microphone. It offers features like private communication, seamless integration, and a discreet, edge-mounted presence. Key features: - Private communication - Seamless integration with Claude Code - Discreet, edge-mounted presence Tech stack: Python Why it matters: Void simplifies the interaction with Claude Code by providing a voice-first interface, making it accessible and convenient for users. Repository: PLAYER99-void/void Primary language: Python GitHub stars: 46 GitHub description: A voice-first desktop companion for Claude Code. Say "hey Void" and a little orb (or a notch in your screen's edge) does the work. Windows and Linux, 100% local voice. README excerpt: A voice-first desktop companion for Claude Code, allowing users to interact with Claude Code through their computer's microphone. It offers features like private communication, seamless integration, and a discreet, edge-mounted presence.

Overview: SEPIA Web-Audio Library is a modular, cross-browser, web audio pipelines framework designed for recording and processing audio in background threads. It includes modules for VAD, ASR, resampling, and more. Key features: - Resampler using Speex codec - Voice-Activity-Detection (VAD) via WebRTC-VAD - Custom VAD module using Meyda - Wave Encoder with lookback-buffer - Porcupine Wake-Word detector - Speech recognition server module Tech stack: JavaScript Why it matters: SEPIA Web-Audio Library simplifies the process of creating modular, background-threaded audio pipelines, making it easier to integrate speech processing and recognition into web applications without blocking the main UI thread. Repository: SEPIA-Framework/sepia-web-audio GitHub stars: 48 GitHub description: A library for recording and processing audio using audio-worklets, web-workers, and script-processors. Primarily focused on speech recording and processing, but can be extended for various use cases. Includes modules for VAD, ASR, resampling, and more. Primary language: JavaScript GitHub description: A library for recording and processing audio using audio-worklets, web-workers, and script-processors. Primarily focused on speech recording and processing, but can be extended for various use cases. Includes modules for VAD, ASR, resampling, and more.

### Overview This repository is the official PyTorch implementation for **SDA** (Synthetic-Domain Alignment), a novel test-time adaptation framework that simultaneously aligns the domains of the source model and target data with the same synthetic domain. The framework is designed to enhance the robustness and performance of models during inference by adapting them to a synthetic domain that closely mimics the target data distribution. ### Key Features - **Synthetic Domain Alignment (SDA)**: SDA is a two-stage test-time adaptation framework that aligns both the source model and target data with a synthetic domain generated from a conditional diffusion model. - **Source Domain Adaptation**: The source-domain model is adapted to a synthetic-domain model through fine-tuning with synthetic data generated from conditional diffusion. - **Target Data Projection**: The target data is projected into the synthetic domain using unconditional diffusion, ensuring domain alignment with the projected target data. - **Unified Adaptation**: SDA is designed to be a unified adaptation framework, making it easier to integrate with existing models and frameworks. ### Tech Stack The repository uses the following technologies: - Python: The primary programming language used for the implementation. - PyTorch: The deep learning framework used for the implementation. - Conditional Diffusion Models: Used for generating synthetic data and aligning the source and target domains. - Unconditional Diffusion Models: Used for projecting target data into the synthetic domain. - ImageNet-C and ImageNet-W: The datasets used for training and evaluation. ### Why It Matters SDA is significant because it addresses the challenge of test-time

### Overview: This repository is a comprehensive code repository for the AI Builders Bootcamp #2, providing foundational knowledge and practical examples in AI development. It covers various stages of AI development, from basic automation to advanced fine-tuning, using Python, OpenAI's API, and other tools. ### Key Features: - **Resume Optimizer**: A project that uses Python, OpenAI's API, and Gradio to build a resume optimization tool. - **Introduction to AI**: Sessions on basic automation scripts, ETL pipelines, and foundational models. - **Software 1.0**: Introduction to AI and building basic automation scripts with Python. - **Software 2.0**: Data engineering, machine learning, and ETL pipelines. - **Software 3.0**: Foundation models, prompt engineering, and fine-tuning. - **Session 4**: Semantic search and text embeddings using RAG. - **Session 5**: Model fine-tuning for LinkedIn post writing and text classification. ### Tech Stack: Python, OpenAI API, Gradio, Plotly Dash, Python libraries for data manipulation and visualization. ### Why it Matters: This repository is essential for developers looking to build robust AI applications, especially in the field of resume optimization and AI job board scraping. It provides a practical, step-by-step guide to implementing AI solutions in various domains, making it a valuable resource for anyone interested in AI development.

### Overview: OmniVDiff is a unified video diffusion framework that enables controllable video generation and understanding. It provides a comprehensive solution for generating videos with specific conditions and interpreting the content of videos. ### Key features: - **Controllable Video Generation**: OmniVDiff allows users to control the content of videos by applying various diffusion processes. - **Video Understanding**: It helps in understanding the semantics and context of videos by providing insights into different modalities such as color, depth, and segmentation. - **Unified Framework**: The framework combines diffusion techniques with other video processing methods to achieve better results. ### Tech stack: Python ### Why it matters: OmniVDiff is significant because it simplifies the process of generating and understanding videos by providing a unified approach that can be applied across various modalities. This makes it easier for developers and researchers to create more sophisticated video applications and models.

### Overview: Nano Banana Pro API (`gemini-3-pro-image-preview`) is an OpenAI-compatible API that allows developers to generate and edit 4K images with multi-step planning. It offers a flexible pricing model starting at $0.03 per image, with no subscription fees and supports various input parameters. ### Key Features: - **Model ID and Routes:** The API uses the model id `gemini-3-pro-image-preview` for image generation and editing. - **Pricing:** Prices are per delivered image, starting at $0.03 per image for 4K resolution, with a tiered pricing model. - **Request Parameters:** The API accepts parameters such as `model`, `prompt`, `size`, `resolution`, `n`, `image_urls`, and `nsfw_check`. - **Quickstart:** A simple Python script is provided to get started with the API, including example curl and Python requests. ### Tech Stack: - Python ### Why It Matters: - **Flexible Pricing:** Starting at $0.03 per image, Nano Banana Pro offers a cost-effective solution without the need for a monthly subscription. - **OpenAI Compatibility:** The API is compatible with OpenAI's API, allowing seamless integration with existing OpenAI models. - **Multi-Step Planning:** The API supports multi-step planning, making it suitable for complex image editing tasks. - **Easy Integration:** The API is designed to be easily integrated into existing projects, making it a valuable tool for developers.

### Overview: This repository contains 8 tested recipes for Qwen Image 2.0 prompts, including examples of renders, aspect ratios, and the cost per image on the APIMart route. The price is $0.02 per image, estimated for the entire set to be $0.16. ### Key Features: - **8 Recipes**: Detailed prompts for Qwen Image 2.0. - **Aspect Ratios**: 1:1, 4:3, 3:2, etc. - **Cost**: $0.02 per image, estimated $0.16 for the set. ### Tech Stack: Python ### Why It Matters: The repository provides a practical guide for developers and designers looking to use Qwen Image 2.0 effectively, ensuring cost efficiency and accurate rendering.

### Overview: This repository contains 8 tested recipes for Gemini Nano Banana 2.1 prompts, including example renders, aspect ratios, and the cost per image on the APIMart route. ### Key Features: - **8 Recipes**: Detailed prompts for various subjects, including food photography, e-commerce, and more. - **Aspect Ratios**: Specific ratios for different image formats (e.g., 1:1 for a square, 4:3 for a widescreen). - **Costs**: The cost per image on the APIMart route, estimated at $0.0137 per image. ### Tech Stack: - Python ### Why It Matters: - **Efficient Pricing**: The cost is $0.0137 per image, making it a cost-effective option. - **Model Compatibility**: Compatible with OpenAI models, allowing seamless integration. - **Flexible Usage**: No need for monthly subscriptions or additional fees; simply pay as you go. - **Asynchronous Nature**: Tasks are processed asynchronously, allowing for easy monitoring and tracking. ### Primary Language: - Python ### GitHub Stars: - 47 ### GitHub Description: - Gemini Nano Banana 2.1 prompts: tested image prompt recipes with example renders, aspect ratios, the model id (`gemini-nano-banana-2.1`) and the cost per image.

### Overview: DiffDoctor is a project that aims to diagnose and treat image diffusion models. It provides a comprehensive solution for identifying and addressing issues in diffusion models, ensuring their effectiveness and reliability. ### Key Features: 1. **Artifact Detector**: A tool for diagnosing artifacts in diffusion models. 2. **Training the Diffusion Model**: A method to train diffusion models using the artifact detector. 3. **Code and Checkpoints**: Available for both training and inference, including checkpoints and code repositories. ### Tech Stack: - Python - Hugging Face Transformers ### Why It Matters: DiffDoctor offers a practical approach to diagnosing and treating diffusion models, ensuring that these models perform optimally and accurately. By providing a solution for artifact detection, it helps in maintaining the integrity and effectiveness of diffusion models in various applications. ### Repository: - Repository: ali-vilab/DiffDoctor - GitHub Stars: 48 - GitHub Description: [ICCV 2025] DiffDoctor: Diagnosing Image Diffusion Models Before Treating ### Installation: - Install using `pip`: `pip install -r requirements.txt` - Download checkpoints using `git lfs` and place them in `checkpoints/` ### Diagnose: - Run `python ad_inference.py` to use the artifact detector demo code. ### Treat: - Modify config/settings.py to change checkpoint paths and training parameters. - Further modify config/settings.py for additional customization. ### Acknowledgements: - This project is based on the codebase of [AlignProp](https://github.com/m

### Overview: Veo 3.1 Lite API (`veo-3.1-lite`) is a cost-effective solution for accessing the Google Veo tier, specifically for text-to-video and image-to-video conversions. It offers a single API key to access 300+ models, including 3.1 Lite, with a pricing model based on generated seconds at various resolutions. ### Key Features: - **Single API Key**: Access to the entire catalog via a single API key, including 3.1 Lite. - **No Subscriptions**: No need for monthly subscriptions or pre-paid plans. - **Immediate Usage**: Charges are applied only after a task is completed, ensuring cost transparency. - **Non-Synchronous Design**: Polls `GET /v1/tasks/{task_id}` for task status updates. ### Tech Stack: Python ### Why It Matters: - **Cost Efficiency**: Ideal for teams looking to save on costs while maintaining flexibility. - **Model Accessibility**: Provides access to 300+ models, including 3.1 Lite, without the need for expensive subscriptions. - **Real-Time Pricing**: Charges are reported in the `cost` field of completed tasks, ensuring accurate billing. - **Simple Integration**: Easy to integrate with existing systems, as it uses the same base URL and `Authorization` header. ### GitHub Stars: 30 ### Primary Language: Python ### GitHub Description: Veo 3.1 Lite API (`veo-3.1-lite`): model id, pricing, parameters and copy-paste examples for the OpenAI