Projects

Overview: Evret is a lightweight framework designed for evaluating Information Retrieval systems, providing standard metrics such as Hit Rate, Recall, Precision, MRR, NDCG, ERR, RBP, and Average Precision. Key features: - Standard Information Retrieval metrics - Token overlap, semantic, and LLM judges for text relevance - Native adapters for Qdrant and other vector databases - Integration with LangChain and LlamaIndex Tech stack: Python Why it matters: Evret simplifies the evaluation of retrieval systems, making it easier to assess the performance of different models and improve the overall quality of Information Retrieval pipelines.

Overview: Qwen-Image 3.0 Pro is an advanced image generation service provided by APIMart, offering high-fidelity images at $0.0285712 per image. It supports multiple resolutions and is compatible with OpenAI's API. Key features: - Supports 1K and 2K output - Pricing: $0.0285712 per image, with a minimum of $1 for the first 1000 calls - Compatible with OpenAI's API - Asynchronous processing with task ID tracking Tech stack: Python Why it matters: Qwen-Image 3.0 Pro provides a cost-effective and scalable solution for generating high-quality images, making it ideal for developers and content creators looking to enhance their projects with professional-grade visuals without the need for a long-term subscription. GitHub stars: 10

### Overview: Pany is a PostgreSQL-native semantic search engine that integrates AI-powered search capabilities directly into your existing database. It allows you to add AI search without the need for separate vector databases, vendor fees, or complex setup. ### Key Features: - **Semantic Search**: Search documents and images using natural language queries. - **Multi-Modal Capabilities**: Supports text and image searches. - **Native SQL Joins**: Seamlessly integrate search results with your existing PostgreSQL data. - **10-Minute Docker Deployment**: Easy setup with Docker, no need for manual configuration. ### Tech Stack: - **Python**: The primary language for development. - **PostgreSQL**: The database engine with the pgvector extension for semantic search. - **pgvector**: Extension for PostgreSQL to handle vectorized queries. ### Why It Matters: - **Integration**: Seamlessly integrates AI search capabilities into your existing PostgreSQL database. - **Cost Efficiency**: No need for separate databases, vector databases, or complex setup. - **Scalability**: Built for performance and scalability, making it suitable for large-scale applications. - **Ease of Use**: Simple installation and configuration, making it accessible for developers. ### GitHub Stars: 19 ### Contact: For further assistance, contact [laxmansrivastacc@gmail.com](mailto:laxmansrivastacc@gmail.com). ### API Endpoints: - **Upload Content**: `curl -X POST http://localhost:8000/embeddings` to upload documents. - **Search**: `curl -X POST http://localhost:8000

### Overview: RaBitQ-rs is a Rust implementation of advanced vector quantization and search algorithms, specifically RaBitQ and MSTG (Multi-Scale Tree Graph). It offers up to 32x memory compression with significantly higher accuracy than traditional PQ or SQ, making it highly effective for approximate nearest neighbor search. ### Key features: - **RaBitQ quantization with IVF search**: Production-ready, achieving up to 32x memory reduction with configurable compression ratios. - **MSTG (Multi-Scale Tree Graph) index**: High performance, enabling production-ready large-scale vector search. ### Tech stack: Rust ### Why it matters: RaBitQ-rs delivers exceptional compression-accuracy trade-offs for approximate nearest neighbor search, offering up to 32x memory reduction with configurable compression ratios, significantly higher accuracy, and adaptive multi-scale clustering for better recall. The library is feature-complete, optimized for fast index building, distance computations, and seamless integration with NumPy and ann-benchmarks.

Overview: Este curso é uma introdução ao engenharia e customização de modelos de linguagem, abordado em português. Ele é construído de baixo para cima, com laboratórios que permitem a mensuração de hipóteses. Key features: - Laboratórios de aprendizado de máquina - Treinamento supervisionado - Avaliação com base em baselines - Sistema de consulta de documentos, ferramentas e limites de custo - Projeto final Tech stack: Python Why it matters: Este curso oferece uma abordagem prática para entender, customizar e avaliar modelos de linguagem, com um foco em engenharia e produção. Ele é escrito em português e pode ser usado por iniciantes ou profissionais, com laboratórios que permitem a mensuração de hipóteses. Repository: lucianoon/llm-course Primary language: Python GitHub stars: 19 GitHub description: Open course on LLM engineering and customization, in Portuguese, from scratch to production, with reproducible labs.

### Overview: 欢迎来到Nano-banana精选图片库！在这里，您可以发现Nano-banana在各种任务场景下生成的令人惊叹的图片和提示词，全方位展示Google在图像生成与编辑方面的无限可能。希望通过这些案例，帮助您更好地了解Nano-banana的强大功能。 ### Key Features: - 多视角图像生成 - 自动修图 - 手绘图控制多角色姿态 - 跨视角图像生成 - 定制人物贴纸 - 动漫转真人Cosplay - 根据地地图像生成 - 真实世界的AR信息化 - 根据食材做菜 - 知识推理生成图像 - 红笔批注 - 爆炸的食物 - 制作漫画书 - 动作人偶 - 地图生成等距建筑 - 参考图控制人物表情 - 插画绘画过程四格 - 虚拟试妆 - 妆面分析 - Google地图视角下的中土世界 - 印刷插画生成 - 超多人物姿势生成 - 物品包装生成 - 疠加滤镜/材质 - 控制人物脸型 ### Tech Stack: 使用了多种技术栈，包括但不限于Python、JavaScript、AI模型等。 ### Why It Matters: 这些案例展示了Nano-banana在图像生成与编辑领域的强大能力，为开发者和创意人员提供了宝贵的

### Overview: Semafold is a vector compression toolkit designed for AI workloads that compresses embeddings, retrieval representations, and KV-cache tensors. It offers 10x compression efficiency, pure NumPy core functionality, and optional GPU acceleration via PyTorch (CUDA/MPS) or MLX (Metal) for professional performance. ### Key Features: - **Compression Efficiency**: 10x compression ratio. - **Pure NumPy Core**: No GPU required, runs anywhere. - **Typed Encode/Decode Contracts**: Ensures clear and explicit data contracts. - **Measurable Byte Accounting**: Provides explicit guarantees on artifact size and distortion. - **Enterprise GPU Acceleration**: Zero-config, automatically offloads to PyTorch (CUDA/MPS) or MLX (Metal) for professional performance. - **Enterprise GPU Acceleration**: Zero-config, automatically offloads to PyTorch (CUDA/MPS) or MLX (Metal) for professional performance. ### Tech Stack: - Python 3.10+ (required for enterprise GPU acceleration). ### Why It Matters: Semafold is essential for teams building AI infrastructure that need measurable storage reduction without sacrificing visibility into distortion, artifact size, or integration boundaries. It ensures that AI models and data remain transparent and maintainable, making it a crucial tool for modern AI development and deployment. ### Repository: - **GitHub Repository**: mindtro/semafold - **GitHub Stars**: 19 - **GitHub Description**: Vector compression toolkit for embeddings, retrieval, and KV-cache with 10x compression efficiency, pure NumPy

### Overview: eyeroll is a Claude Code plugin designed to analyze video footage, screen recordings, and Loom videos. It helps coding agents fix bugs, build features, and create skills by providing structured summaries and recommendations based on the video content. ### Key Features: 1. **Video Analysis**: Identifies key elements in video footage, such as bugs, features, or questions. 2. **Codebase Context**: Provides context for the video, including intent, category, and confidence levels. 3. **Bug Fixing**: Automatically diagnoses and fixes bugs in the codebase. 4. **Skill Creation**: Converts video tutorials into skills by analyzing the content. 5. **Command Line Interface**: Offers a command-line interface for quick analysis and bug fixing. ### Tech Stack: The project is built using Python and relies on various libraries and APIs for video analysis and bug fixing. It supports multiple backend options including Gemini, TwelveLabs, OpenAI, and Ollama. ### Why It Matters: eyeroll streamlines the process of analyzing video content for coding agents, making it easier to fix bugs, build features, and create skills. The structured summaries and context provided by the plugin enhance the efficiency and effectiveness of the coding process. ### Repository: The repository is hosted on GitHub at [mnvsk97/eyeroll](https://github.com/mnvsk97/eyeroll). It has been starred by 19 users, indicating its usefulness and popularity among developers. ### Installation: To use eyeroll, you can add it to Claude Code using the following commands:

### Overview: This project provides a tutorial for developers on how to design, deploy, and evaluate large language models (LLMs) for inference. It focuses on creating a C/S style optimized LLM inference engine, which is optimized for single GPU usage. The tutorial includes detailed steps on setting up the environment, deploying services, and evaluating performance across different traffic characteristics. ### Key Features: - **Optimized C/S Style Inference Engine**: A high-performance LLM inference engine that is designed to be deployed on a single GPU. - **Static Batching (SB) and Continuous Batching (CB) Strategies**: Support for both static and continuous batch processing strategies. - **xformers and PagedAttention Acceleration**: Utilizes xformers and PagedAttention to speed up inference. - **GPU-CPU Memory Exchange**: Efficient memory management between GPU and CPU. - **Static GPTQ Model Loading**: Supports loading `int4` GPTQ models. - **Client Load Balancing**: Provides client-side load balancing for better performance. - **Static and Dynamic Batch Processing**: Supports both static and dynamic batch processing strategies. - **Support for llama v1 and llama v2 (except 70B)**: Models supported by the engine. - **Support for `safetensors` Weight Files**: Only supports `safetensors` format for weights. - **No Support for Extending Context Length**: Limited to a fixed maximum context length. - **No Support for Adapter(s)**: No support for adapter modules. - **No Support for Truncation Sampling

Overview: pyobvector is a Python SDK designed for OceanBase Multimodal Store, providing vector search, full-text search, and JSON table operations. It supports both Milvus-compatible API and SQLAlchemy-based SQL mode, offering both OceanBase and OceanBase seekdb support. Key features: - Milvus-compatible API - SQLAlchemy-based SQL mode - Embedded SeekDB mode - Hybrid Search mode Tech stack: Python Why it matters: pyobvector simplifies the integration of OceanBase Multimodal Store with Python applications, enabling efficient vector search, full-text search, and JSON table operations, making it a valuable tool for developers working with OceanBase databases.

Overview: This project is a Docker build of PostgreSQL 15.3 with the cube extension, specifically raising the dimension limit to 350, which is necessary for working with machine learning related vectors. Key features: - Docker build of PostgreSQL 15.3 - Cube extension with dimension limit raised to 350 Tech stack: Dockerfile Why it matters: The cube extension in PostgreSQL has a hard limit of 100 dimensions per vector, which is problematic for machine learning applications. By raising this limit to 350, this project allows for better compatibility with vectors used in machine learning, particularly for tasks like word embedding. GitHub stars: 19

### Overview: Fine-tuning whisper models on Vietnamese using Jupyter Notebooks, leveraging Kaggle, Colab, and AWS EC2. This project includes traditional and PEFT LoRA fine-tuning approaches, with models evaluated for Word Error Rate (WER). ### Key Features: - Traditional fine-tuning of whisper models. - Fine-tuning whisper models with PEFT-LoRA and INT8. - Fine-tuning wav2vec BERT v2 models. - Evaluation of whisper models using batched inference. - Conversion to OpenAI Whisper, whisper.cpp, and ONNX formats. ### Tech Stack: Jupyter Notebook ### Why it Matters: This project provides a comprehensive guide for fine-tuning whisper models on Vietnamese, including traditional and advanced fine-tuning techniques, and offers evaluation metrics for model performance.

Overview: PomaiDB is a lightweight, embeddable vector database and multimodal AI memory engine designed for low performance devices, such as Raspberry Pi, Orange Pi, NVIDIA Jetson, and other edge hardware. Key features: - Vector search with configurable ANN indexes - Offline-first RAG pipeline for local AI inference and memory - Multimodal storage and key-value storage Tech stack: C++ Why it matters: PomaiDB prioritizes predictable latency, low memory usage, and sequential storage I/O, making it suitable for edge devices with limited resources, such as cameras, gateways, and custom operating systems. Repository: pomagrenate/pomaidb GitHub stars: 19 GitHub description: Embedded Vector Database for low performance devices

### Overview: This repository contains a collection of text-to-image generation gems and libraries, designed to create unique and pixel art avatars from text attributes. The projects include various characters such as owls, cyberpunks, beaux & belles, cats, shiba inus, goblins, and more, with options for customization and zooming. ### Key Features: - **moonbirds**: Generate 42x42 pixel owl images from text attributes. - **cyberpunks**: Create 32x32 pixel cyberpunk avatars from text attributes. - **nouns**: Generate 32x32 pixel avatar images from text attributes. - **beaux & belles**: Create 24x24 beau & belle head avatars from text attributes. - **coolcats**: Generate 24x24 pixel cat images from text attributes. - **shiba-inus**: Create 24x24 pixel shiba inu doge images from text attributes. - **goblins & lil'goblins**: Generate 24x24 goblins or 20x20 lil'goblins images from text attributes. - **readymades**: Generate 24x24 pixel punk images from 1/1 originals. - **smileys**: Create 24x24 pixel smiley images from text attributes. ### Tech Stack: Ruby ### Why It Matters: These projects allow users to create unique and personalized pixel art avatars from text descriptions, making them perfect for creative projects,

Overview: KoutenDB is a locality-aware NoSQL document/vector database designed for reducing the amount of data an application must process and reduce the memory and token usage in retrieval operations. Key features: - Locality-first document and vector database - Reduces records, bytes, candidate memory, and tokens processed by an application - Supports RAG and AI retrieval, as well as tenant-scoped web systems, user detail pages, product data, application state, and any service where related data is known before a read begins - Uses `ring` coordinates for efficient data placement and retrieval - Supports subrings for more granular control over data retrieval - Meets strong-durability requirements with zero client errors during a 72-hour local three-node, disk-backed, strong-durability run Tech stack: Nim Why it matters: KoutenDB simplifies data management by reducing the amount of data an application must process and the memory and token usage in retrieval operations, making it ideal for RAG and AI retrieval, as well as tenant-scoped web systems, user detail pages, product data, application state, and any service where related data is known before a read begins.

Overview: Qdrant VSQL is a Python library designed to simplify vector database querying by translating SQL-like `WHERE` clauses into Qdrant's `Filter` objects. It aims to make complex filtering logic more intuitive and efficient. Key features: - SQL-like syntax for `WHERE` clauses - Comprehensive operator support including `=`, `!=`, `<>`, `>`, `>=`, `<`, `<=`, `IN`, `NOT IN`, `BETWEEN`, `NOT BETWEEN`, `LIKE`, `IS NULL`, `IS NOT NULL`, `IS EMPTY`, `COUNT()` - Logical operators `AND`, `OR`, `NOT` - Nested field support for dot notation and array projection - Type handling for string, number, and boolean values - Qdrant native output for efficient filtering - Automatic translation of `id` field operations - Case-insensitive handling for operators and keywords - Full-text search capabilities with `LIKE` operator Tech stack: Python Why it matters: Qdrant VSQL simplifies the process of querying vector databases by making complex filtering logic more accessible and efficient, potentially becoming a standard for vector database interactions.

### Overview: AFM Trainer is a comprehensive GUI wrapper application for Apple's Foundation Models Adapter Training Toolkit, designed to provide an intuitive interface for training LoRA adapters on Apple's on-device foundation models. ### Key Features: - **Modern GUI**: Beautiful, professional interface with dark/light theme switching. - **Enhanced Theming**: Modern Sun Valley theme for a polished, forest-inspired appearance. - **Real-time Monitoring**: Live training progress with loss visualization and log streaming. - **Complete Workflow**: From dataset validation to .fmadapter export. - **Flexible Configuration**: All Apple toolkit parameters exposed with sensible defaults. - **WandB Integration**: Optional Weights & Biases logging for advanced metrics tracking. - **Comprehensive Error Handling**: User-friendly error messages and crash reporting. - **UV Environment Management**: Transparent dependency handling and installation. - **Smart File Management**: Automatic .gitignore updates and dataset validation. - **Draft Model Training**: Optional speculative decoding support. - **Configuration Profiles**: Save and load training configurations. - **Theme Customization**: Real-time switching between dark and light themes. ### Tech Stack: Python 3.11+ ### Why it Matters: AFM Trainer simplifies the process of training LoRA adapters for Apple's on-device foundation models, making it accessible and user-friendly for developers.

Overview: Train JEV-style System 1 decision models on Qwen3.5 and Qwen3.8-27B with a 24-slot decision head, KL distillation, calibration, and serving capabilities. Key features: - Decision models trained on Qwen3.5 and Qwen3.8-27B - 24-slot decision head - KL distillation for improved performance - Calibration for better accuracy - Serving API compatible with TypeSafe System One API Tech stack: Python Why it matters: This project allows for the training and deployment of JEV-style System 1 decision models, leveraging Qwen3.5 and Qwen3.8-27B with a 24-slot decision head, KL distillation, and calibration to enhance model performance and reliability. The decision models are designed to be easily integrated into existing applications, making them a valuable tool for decision-making processes.

Overview: A high-performance GEMM implementation targeting NVIDIA RTX 50 series (SM120) GPUs, achieving 100% cuBLAS performance with no dependencies beyond the CUDA toolkit. Key features: - 100% cuBLAS performance - TMA (Tensor Memory Accelerator) optimization - Warp Specialization for efficient memory and compute overlap - Multi-stage Software Pipeline - Persistent Kernel - Tensor Core via `mma.sync.aligned.m16n8k16` - Split-K for small-M problems - 128B Shared Memory Swizzling - Swizzled Tile Rasterization - `ldmatrix` / `stmatrix` for efficient MMA fragment movement - L2-aware Benchmarking - JIT Kernel Compilation - Autotuning Tech stack: CUDA C++ with inline PTX Why it matters: This library offers a zero-dependency GEMM solution for RTX 50 series GPUs, achieving 100% cuBLAS performance, making it highly efficient and suitable for applications requiring high performance on these specific hardware. GitHub stars: 19

Overview: A powerful RAG system that leverages real documents for more reliable AI insights. Key features: - Integrates with OpenAI GPT-4o for advanced language understanding. - Utilizes LangChain for AI logic. - Stores document text in ChromaDB for efficient search. Tech stack: LangChain, OpenAI GPT-4o, ChromaDB. Why it matters: Provides more accurate and reliable AI insights by grounding it in real documents, reducing the risk of hallucinations.