Projects

Overview: This project is a Model Context Protocol (MCP) server designed to integrate with Turso databases for LLMs, providing easy management and querying of Turso databases from LLMs. Key features: - List Databases - Create Database - Delete Database - Generate Database Token - List Tables - Execute Read-Only Query - Execute Query - Describe Table - Vector Search Tech stack: TypeScript Why it matters: This server implements a sophisticated two-level authentication system, ensuring secure and efficient management of Turso databases from LLMs, with a focus on read-only and read-write operations. Repository: spences10/mcp-turso-cloud Primary language: TypeScript GitHub stars: 18 GitHub description: A Model Context Protocol (MCP) server that provides integration with Turso databases for LLMs. This server implements a two-level authentication system to handle both organization-level and database-level operations, making it easy to manage and query Turso databases directly from LLMs.

### Overview: USAF is a fine-tuning method designed for hardware with limited resources, specifically a 12GB GPU. It trains MoE models on 0.5% of the expert parameters, significantly reducing the required VRAM and computational resources. ### Key Features: - **Trains Expert Weights and Router**: USAF combines the training of expert weights and the router, which is essential for MoE models. - **Runs on 12GB GPU**: USAF can run on a 12GB GPU, making it accessible to hardware with limited resources. - **Trains 0.5% of Expert Parameters**: It trains 0.5% of the expert parameters, significantly reducing the VRAM requirements compared to full fine-tuning. ### Tech Stack: Python ### Why It Matters: USAF is a method that can be used to fine-tune MoE models on hardware that cannot support full fine-tuning, making it accessible to developers and researchers with limited computational resources.

### Overview: `vindex` is an open-source video knowledge compiler that turns videos into structured knowledge artifacts for AI agents. It produces stable, typed, versioned output artifacts in JSON and Markdown, enabling AI systems to consume video understanding without reprocessing the raw video. ### Key features: - **Stable, Typed, Versioned Output**: Produces structured JSON and human-readable Markdown artifacts. - **Versatile Extraction Pipeline**: Modular multi-modal extraction pipeline for videos. - **Deterministic Compilation**: Produces facts, not opinions, and runs entirely on your machine without cloud account, API key, or data leaving your device. - **Versatile Use Cases**: Suitable for AI coding agents, video editing automation, content search and indexing, lecture analysis, and RAG pipelines over video libraries. ### Tech stack: Python ### Why it matters: `vindex` addresses the problem of rederiving video understanding from scratch, which is a common challenge faced by AI systems. By compiling video understanding into stable, versioned JSON and Markdown artifacts, `vindex` enables AI systems to consume video knowledge immediately, without the need for reprocessing the raw video, and without the need for cloud-based services or database servers. This makes `vindex` a valuable tool for developers and researchers looking to improve the efficiency and reliability of video-based AI systems.

Overview: vLLM Doctor is a tool designed to diagnose bottlenecks in vLLM inference servers by analyzing live metrics. Key features: - Root-cause summary: A clear diagnosis of the most probable cause, including evidence behind it and recommended next steps. - Built-in diagnosis rules: Identifies common bottlenecks such as queue pressure, TTFT/TPOT bottlenecks, KV cache pressure, low throughput, error rate, replica imbalance, prefix cache efficiency, preemption pressure, and queue latency. - Local history: Stores and reviews past runs with `--save` and `history list`/`history show`. - Watch change-log: Only persists state changes (health or firing rules) with `--save --watch`. - Dual input: Supports Prometheus or direct `/metrics` scrape. - Structured output: Rich text tables or JSON for automation. - Configurable thresholds: Allows tuning of rules per rule via TOML. Tech stack: Rust. Why it matters: vLLM Doctor provides a fast, server-side diagnostic snapshot for a single vLLM server or Prometheus target, offering a clear diagnosis of bottlenecks without requiring setup or dashboard replacement. Repository: vllm-doctor/vllm-doctor GitHub stars: 18 GitHub description: Diagnose vLLM inference servers. README excerpt: [Image of vLLM Doctor's wordmark](https://raw.githubusercontent.com/vllm-doctor/vllm-doctor/main/docs/assets/wordmark.svg) ```shell vllm-doctor diagnose http://localhost:

Overview: Seedream Image MCP is a powerful tool that leverages the SeeDream model from VolcEngine to generate high-quality images. It supports customization of dimensions and smart reference images, making it easy to integrate AI-generated content into various applications. Key features: - Uses the SeeDream 4.0 model for high-quality image generation - Supports custom dimensions and smart reference images - Automatically generates image prompts based on user needs - Supports MCP protocol, allowing seamless integration with Cursor and Claude Desktop Tech stack: TypeScript Why it matters: Seedream Image MCP simplifies the process of integrating AI-generated content into applications, making it more accessible and efficient for developers. Repository: wearzdk/seedream-image-mcp GitHub stars: 18 GitHub description: 🚀 PixelMCP | 为你的 Cursor、Claude Code 等集成AI绘画能力，让AI生成的页面不再单调！ README excerpt: # SeeDream Image MCP 基于火山引擎 SeeDream 模型的 MCP (Model Context Protocol) 图片生成工具。 ## ✨ 特性 - 🎨 使用火山引擎 SeeDream 4.0 模型生成高质量图片 - 🔧 支持自定义尺寸、智能参考图等 - 📝 无需编写复杂提示词，AI自动根据需求生成生图提示词 - 🔌 MCP 协议支持，可在 Cursor、Claude Desktop 等客户端中使用 ## 📺 演示 <video src="https://github.com

Overview: This repository provides end-to-end scripts demonstrating various features and integrations with Weaviate, a machine learning model hosting service. It serves as a reference for JavaScript developers on how to use the Weaviate client. Key features: - Similarity Search - Generative Search Tech stack: TypeScript Why it matters: This repository offers practical examples of how to utilize Weaviate's features, making it a valuable resource for developers looking to integrate machine learning models into their projects. Repository: weaviate/recipes-ts GitHub stars: 18

### Overview: This repository contains the code used to fine-tune a model trained with Sparse Autoencoders (SAE) on CLIP, aiming to improve the adversarial training of CLIP models. The project is experimental and serves as a fun exploration into the intersection of SAEs and CLIP fine-tuning. ### Key features: - **SAE-Informed Adversarial Training**: Fine-tuning a model trained with SAEs on CLIP to improve adversarial training. - **Adversarial Typographic Attack Dataset**: Curated using SAE to test the saliency of text in CLIP models. - **Model Conversion**: Converts the GmP model back to a normal CLIP model for downstream tasks. - **Datasets**: Contains datasets required for adversarial training and zero-shot testing. ### Tech stack: Python ### Why it matters: This project showcases the potential of combining SAEs with CLIP for adversarial training, providing insights into how SAEs can influence the learned representations of CLIP models. The results highlight the importance of understanding the underlying mechanisms of SAEs and their impact on CLIP's adversarial training. ### GitHub stars: 18

### Overview: This repository contains code for multi-scale fine-tuning on LSF tasks, designed to improve the performance of time series foundation models for forecasting. The approach leverages disentangled scale effects to enhance model accuracy. ### Key Features: - Multi-scale fine-tuning - Built on top of Moirai (Uni2TS codebase) - Fine-tuning setup follows Moirai's configuration - Provides code for LSF tasks - Disentangles scale effects to better leverage time series foundation models ### Tech Stack: Python ### Why It Matters: This method aims to improve the accuracy of time series forecasting models by effectively handling scale effects, making it particularly useful for applications where scale variations significantly impact model performance.

**Overview:** PersonOS is an open-source framework designed for long-term memory management in personal agents, organizing preferences, relationships, goals, and experiences into retrievable memories across conversations. **Key features:** - **Memory organized around people** - Extract personal information from conversations, retain its context, and scope memories by user and session. - **Recall with evidence** - Retrieve relevant memories and their sources as material for your agent, or generate answers with citations. - **Context that keeps its history** - Connect new and earlier statements through memory chains, retain past records, and consolidate episodes into user profiles. - **Composable integration** - Use Python or HTTP, configure models, storage and retrieval behavior, and enable multimodal backends as needed. **Tech stack:** Python **Why it matters:** PersonOS is essential for personal agents that need to remember and recall information over extended periods, providing a robust and context-aware memory management system. **GitHub stars:** 16 **Repository:** 2049lab/PersonOS

### Overview: SCoRe: Training Language Models to Self-Correct via Reinforcement Learning SCoRe is a task-agnostic implementation of the paper "Training Language Models to Self-Correct via Reinforcement Learning" (Kumar et al., ICLR 2025), built on the Unsloth LoRA architecture for small open-weight models. The core algorithm is identical to the original paper, with adaptations that improve performance and ease of use. ### Key Features: - **Reasoning Tagging**: Models reason inside `<think>...</think>` and give the final answer inside `<answer>...</answer>`. This format is scored directly. - **K3 KL Estimator**: Unbiased forward KL from per-token log-probs only, avoiding `[B, T, V]` log-softmax tensor. - **Optional Length Normalization**: The default `"sequence"` divides the PG and K3 KL terms by the actual per-sample generated length. The `"constant"` option divides by `max_new_tokens`, avoiding length bias. - **Compound Reward**: 0.25 for a `<think>...</think>` pair, 0.25 for one `<answer>...</answer>` pair, 0.5 for extracted-answer match. This gives the α-bonus more signal. - **LoRA-only**: Reference policy = same model with `model.disable_adapter()`. No second model in VRAM. - **Optional Stage II**: Adds the reward-shaping bonus `α · (r(y2) − r(y1))` for improved performance. ### Tech

### Overview: Kevrai Omni 是一个一键本地 AI 工作站，提供包括 LLM、TTS、图像、视频、3D、超分、音乐、音源分离、ASR、语音合成、知识库与超级搜索等多功能。它使用 Electron + 原生 HTML/CSS/JS 构建，支持本地模型管理、硬件检测与推荐、多引擎支持（llama.cpp、MNN、vLLM、ComfyUI）以及数据本地化存储。 ### Key features: - **本地模型管理**: 支持从硬盘导入 GGUF、safetensors 文件或整个文件夹，直接拖入窗口。 - **模型市场**: 模型浏览、搜索、筛选，支持模糊匹配、拼音式缩写、中文搜索。 - **硬件检测与推荐**: 自动检测 CPU、内存、磁盘、GPU（nvidia-smi、rocm-smi、npu-smi、system_profiler），推荐合适的模型与量化档。 - **模型收藏与最近使用**: 收藏模型、记录最近打开的模型，可在工具栏筛选。 - **本地模型目录**: 列出所有可用 GGUF 量化文件，支持单独下载。 - **AI 引擎管理**: llama.cpp 等引擎按需下载、安装、卸载、检查更新，已装自动跳过。 - **MNN 推理**: 提供模型市场、下载、加载与对话。 - **LLM 对话**: 支持流

### Overview: Fine-tuning Large Language Models on Consumer-Grade Hardware This project implements a resource-efficient method for fine-tuning Large Language Models (LLMs) on consumer-grade hardware. It focuses on the technical implementation of Parameter-Efficient Fine-Tuning (PEFT), offering a modular codebase that supports two distinct training pathways: the standard Hugging Face library and the optimized Unsloth framework. ### Key Features: - **Parameter-Efficient Fine-Tuning (PEFT)**: Utilizes Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA) to fine-tune models on low-resource hardware. - **LoRA**: Freezes the pre-trained model and injects trainable rank-decomposition matrices into transformer layers, reducing computational and memory requirements. - **QLoRA**: Quantizes the frozen base model to 4-bit precision, significantly reducing memory usage while maintaining performance. - **Unsloth Mode**: Optimized low-level GPU kernels for faster and more efficient QLoRA fine-tuning, up to 5x faster than standard implementations. - **Resource-Efficient**: Designed for low-resource environments, making it suitable for fine-tuning billion-parameter models on standard GPUs. ### Tech Stack: - **Python**: The primary programming language used for the implementation. - **PyTorch**: Used for the fine-tuning process. - **CUDA**: Optimized for GPU execution, providing superior VRAM utilization. - **Unsloth**: Framework for optimized low-level GPU kernels. - **Hugging Face**: Library for fine-tuning models

### Overview: Fusion Embedding is an open-weight, self-hostable model that maps seven modalities (text, image, video, audio, thermal, inertial motion, and touch) into a single shared vector space. It extends a state-of-the-art vision-language base with new senses, preserving the base's performance and being fully self-hostable. ### Key features: - **Unified Vector Space**: Maps text, image, video, audio, thermal, inertial motion, and touch into a single shared vector space. - **Freezing Base**: Only audio-side components and a temperature are trained, with the base's MMEB-V2 performance inherited unchanged. - **Matryoshka Embeddings**: Truncates to any rung of 2048, 1536, 1024, 512, 256, 128, or 64, preserving the base's performance. - **Self-Hostable**: The model can be deployed and run on any device or platform without modification. ### Tech stack: Python ### Why it matters: Fusion Embedding is designed to be a versatile model that can be used for various applications, including retrieval, RAG, clustering, and cross-modal search. Its ability to map multiple modalities into a single vector space and its self-hostable nature make it a valuable tool in the developer's toolkit.

### Overview: 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: - **PDF upload**: Multipart upload with server-side parsing, chunking, and vector indexing. - **Question answering**: Semantic search over document chunks, then local HF model generation. - **Reading Modes**: Choose between Standard, Tutor, Socratic, Simple, or Concise answering styles. - **Summarization**: Bullet-style summaries from retrieved context. - **Multi-document UI**: Upload and switch between multiple PDFs. - **In-browser viewer**: Page-by-page PDF preview. - **Chat export**: Export conversation history as CSV or plain text. ### Tech stack: JavaScript, React, Node.js, Express, FastAPI, LangChain, Ollama, Hugging Face, Chroma, Streamlit. ### Why it matters: This application streamlines the process of extracting and summarizing information from PDFs, making it easier for users to quickly understand and engage with the content. The integration of RAG technology ensures that the generated answers are contextually relevant and accurate, enhancing the user experience.

### Overview: DeepSeek Harness 插件：DeepSeek Pro 大脑 + 自动识图。GUI 附加图片默认经官方 deepseek-v4-flash-vision-exp 原生识图，转译成文字后交给 DeepSeek 作答（纯文本的 V4-Pro 也能看图）；支持百炼/智谱/OpenRouter 等任意 OpenAI 兼容 VLM，无 key 自动探测本地 Ollama；安装时有一问式确认。 ### Key features: 1. **原生识图直通**：选中模型支持图片时，图片原样透传，不再转译。 2. **内层路由自动发现**：配置的内层路由未注册时，按顺序自动挑选：`deepseek-official` → `deepseek-account` → 任一已注册路由。 3. **零配置本地路径**：启动时探测 `http://localhost:11434`，检测到 Ollama 即自动加入降级链。 4. **快速且明确的失败**：没有 key 也没有本地 Ollama 时，转译失败并给出指引。 5. **有 key 自动提速**：导出 `VISION_API_KEY` / `DASHSCOPE_API_KEY` 后自动走付费端点。 6. **降级链 + 错误分类**：分类给出可操作提示；占位符报告每一次尝试。 7. **思考型模型兜底**：如 `qwen

### Overview: AI Cactus is a lightweight plugin for integrating Cactus AI framework into Unreal Engine 5, allowing developers to run local LLMs at runtime. ### Key Features: - **LLM Features:** - Generate Text: Answers questions without storing conversation history. - Run Conversation: Records questions for future reference. - Clear Conversation: Resets conversation history. - Export and Import Conversation: Saves conversation for persistence. - **VLM Features:** - Response to Image: Answers questions about images without storing conversation history. - Conversation About Image: Records questions for future reference. - Clear Conversation: Resets conversation history. - Export and Import Conversation: Saves conversation for persistence. ### Tech Stack: Unreal Engine C++ ### Why it Matters: AI Cactus simplifies integrating LLMs into Unreal Engine projects, ensuring lightweight performance and easy persistence of conversations.

Overview: IGReg is an image-geometry-assisted point cloud registration method that addresses the challenges in indoor scenes by integrating image information and selectively fusing geometric consistency. Key features: - Intra-Image-Geometry fusion module for integrating texture and structure information into the point feature space. - Selective correlation fusion module for embedding correlations between salient anchors and points. - Iterative information exchange between saliency location and selective correlation fusion modules. Tech stack: Python, PyTorch, CUDA 11.7, PyTorch 1.13.1, torchvision 0.14.1+cu117, torchaudio 0.13.1. Why it matters: IGReg effectively improves point cloud registration performance in indoor scenes by leveraging image information and selectively fusing geometric consistency, leading to accurate correspondence matching and successful registration in geometrically challenging cases such as repetitive patterns and low-geometry regions. GitHub stars: 17, Repository: Jiang0903/IGReg

Overview: TuneForge is an MCP server designed to generate datasets, fine-tune, and evaluate Large Language Models (LLMs) directly from a coding agent. Key features: - Generates SFT datasets from a product description and optional source text, with LLM-judge quality filtering. - Runs LoRA SFT on any Hugging Face causal LM. - Continues with policy-gradient RL using an Ollama-hosted teacher as the judge. - Merges adapters, evaluates models on held-out data, and polls job status. Tech stack: Python Why it matters: Fine-tuning is commoditized, but making the whole loop — data → training → evaluation → merge → redeployment — drivable from inside the agent you already talk to is not. TuneForge is that loop, packaged as an MCP server. GitHub stars: 17 Repository: JustVugg/tuneforge

### Overview: FlashSpec is a lossless speculative decoding library for HuggingFace causal language models, designed for efficient inference on CUDA GPUs. It optimizes the decoding process by leveraging a static-cache CUDA-graph fast path, resulting in significant performance improvements over standard HuggingFace `generate` on a single GPU. ### Key features: - **Lossless Speculation**: FlashSpec ensures the same distribution and tokens as the model's output, providing exact results. - **Static-Cache CUDA-Graph Path**: Utilizes a static-cache path to accelerate decoding, reducing the overhead of dynamic graph compilation. - **Small HTTP Server**: Includes a simple HTTP server for serving the model, making it easy to integrate with other systems. ### Tech stack: FlashSpec is built using Python and PyTorch, utilizing CUDA for GPU acceleration and static caching for performance optimization. ### Why it matters: FlashSpec is particularly useful for developers and researchers working with HuggingFace models, especially when dealing with large language models. It provides a faster and more accurate decoding process, making it ideal for applications requiring high performance and correctness.

### Overview: This project involves fine-tuning GPT-2 for text summarization using Python. The repository includes detailed steps for downloading and preprocessing data, fine-tuning the model, and evaluating the results using ROUGE scores. ### Key features: - Fine-tuning GPT-2 for text summarization - Data preprocessing and splitting - GPT-2 tokenizer and model fine-tuning - Generation of summaries for test data - Evaluation using ROUGE scores ### Tech stack: Python ### Why it matters: This project provides a comprehensive guide for fine-tuning GPT-2 for text summarization, making it useful for developers looking to improve their text summarization capabilities.