Projects

Overview: CTTC is a Python-based Flask application designed to analyze cyber threat research posts from various URLs, leveraging ChatGPT to provide insights such as prevention, detection, and summary. This project aims to streamline the process of understanding and addressing cyber threats. Key features: - Scrapes cyber threat research posts from given URLs - Integrates ChatGPT for context analysis - Provides prevention, detection, and summary insights Tech stack: Python Why it matters: CTTC offers a faster and more efficient way to analyze cyber threat research, especially for those who lack the time or expertise to manually read and interpret threat reports. It streamlines the process and can be used by experienced human analysts to enhance their cyber threat intelligence capabilities. Repository: whichbuffer/CTTC Primary language: Python GitHub stars: 11

### Overview: Moonfrost is a state-of-the-art 777M-parameter Mixture-of-Experts language model trained from scratch, designed for high-performance chat and instruction-following tasks. ### Key features: - **777,148,032 Parameters**: Active for any given token. - **Multi-head Latent Attention**: Compressed key-value cache. - **DeepSeekMoE Feed-Forward Layers**: Selects 3 of 32 experts per layer. - **Fine-tuned for Chat**: Trained on 2.7B tokens of instruction data. - **Training Cost**: Approximately $55, took about 12 H100-hours. ### Tech stack: Python ### Why it matters: Moonfrost is a powerful tool for generating high-quality responses to text-based tasks, making it ideal for chatbots, language models, and more. Repository: whoashish115/moonfrost-ai

Overview: This project provides a graphical user interface (GUI) for GPT4All, allowing users to interact with LLMs in a more intuitive way. Key features: - A self-contained Python script named `appGUI.py` for running the GUI. - Support for various Python bindings and packages, including `gpt4all`, `typer`, and `tkinter`. - Option to specify the model path, prompt template, context length, number of threads, and device for inference. - System prompt and temperature settings available for customization. Tech stack: Python bindings for GPT4All. Why it matters: This tool simplifies the process of running large language models (LLMs) by providing a more user-friendly interface, allowing developers and users to interact with models without needing to delve into complex command-line interfaces.

### Overview: Midjourney Prompts: 8 Tested Recipes with Outputs <!-- conv-kit:v1 --> <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-hero-perfume.jpg" width="820" alt="Sample render via the APIMart image API"> </p> > **from $0.045 per image** on the APIMart route — un único endpoint compatible con OpenAI en `https://api.apimart.ai/v1`, sin plan mensual. *(observado el 2026-10-12)* **[Get an API key](https://go.apimart.ai/k-e7900c)** · **[Live pricing](https://go.apimart.ai/k-bf32d4)** · **[Model page](https://go.apimart.ai/k-b31f09)** · [⚡ 60-second quickstart](#quickstart) **Por qué los equipos usan Midjourney (`midjourney`) a través de APIMart** - **Una clave, todo el catálogo.** La misma base URL `https://api.apimart.ai/v1` y la misma cabecera

Overview: Lillama is a project that aims to compress large language models (LLMs) by distilling low-rank features, specifically targeting the Microsoft Phi-2 model. This approach reduces the model size significantly without compromising performance. Key features: - Distillation of low-rank features from large models - Compression of the Microsoft Phi-2 model by 20% using distillation - Generation of detailed evaluation results using `lm-eval` and manual inspection with Huggingface Transformers - Ability to share the compressed model on Huggingface for community use Tech stack: Python, Flash Attention, Huggingface Transformers, PyTorch Why it matters: This project demonstrates innovative techniques for reducing the size of large language models while maintaining or improving their performance, which is crucial for resource-constrained environments and applications where model size is a significant concern.

### Overview: CodeClaw is a CLI tool designed to convert Claude Code and Codex sessions into privacy-safe training datasets. It includes features such as redaction, background sync, and MCP memory retrieval, making it suitable for structured, reusable training data with privacy controls. ### Key Features: - **Privacy Redaction:** Ensures that sensitive information is redacted. - **MCP Memory Retrieval:** Preserves historical problem-solving context. - **Multi-Source Ingestion:** Supports Claude Code and Codex sessions, as well as experimental adapters for various coding environments. - **Privacy-Aware Export:** Reduces the risk of privacy breaches through multiple layers of privacy protection. - **Controlled Publishing Workflow:** Allows for local export, user review attestations, and publish to Hugging Face with explicit approval. - **Continuous Mode:** Background watch daemon for incremental sync. - **Memory Tooling:** MCP server with features like project patterns, trajectory stats, and graph similarity retrieval. ### Tech Stack: - Python ### Why It Matters: - Streamlines the process of turning coding sessions into structured training data. - Provides privacy controls to protect sensitive information. - Preserves historical context for problem-solving, enhancing the learning process. ### Repository: - Repository: [ychampion/codeclaw](https://github.com/ychampion/codeclaw) - GitHub Stars: 11 - GitHub Description: CLI for exporting Claude Code/Codex sessions to Hugging Face with privacy redaction, MCP memory, and share-ready dataset workflows. ### Installation: - **Install:** ```bash

### Overview: WePost 是一个一站式社交媒体卡片生成器，能够将文字内容一键转换为精美卡片图，适用于小红书、朋友圈、公众号封面、早报、代码截图等场景。用户只需输入文字，即可生成符合多种风格和尺寸的卡片，支持高清导出。 ### Key features: - **卡片模板预览**：提供10套精心设计的卡片模板，涵盖多种风格和尺寸。 - **Markdown支持**：支持标题、段落、引用、列表、表格、代码块等Markdown语法。 - **卡片渲染**：自动匹配模板与画幅，实时渲染为精美卡片。 - **导出功能**：支持高清导出，可直接用于社交媒体平台。 - **生产服务**：提供静态部署和完整生产服务，支持无头渲染接口。 ### Tech stack: - TypeScript - Next.js - TypeScript - MIT License ### Why it matters: WePost简化了内容创作者的创作流程，使其能够快速、高效地将文字内容转化为吸引眼球的卡片图，适用于多种社交媒体平台，提升内容的传播效果。

Overview: 这是一个本地优先的SFT训练数据格式转换工具，将常见的JSON、JSONL和CSV格式统一为OpenAI的`messages`格式，并生成LLaMA-Factory可使用的`train.jsonl`和`val.jsonl`文件。此工具提供了一个简单的网页界面，使得非技术用户可以轻松地处理各种格式的训练数据。 Key features: - 支持JSON、JSONL、CSV格式 - 识别OpenAI `messages`、ShareGPT `conversations`、Alpaca、`prompt/response`、`question/answer`、`query/response`等格式 - 转换前预览实际`messages` - 校验角色顺序、空内容、非文本消息和不完整对话 - 使用固定随机种子，可复现地划分训练集与验证集 - 原子写入文件，转换失败时不会留下半份结果 - 无第三方运行依赖，训练数据不会主动上传到外部服务 - 提供macOS双击启动包的构建脚本 Tech stack: Python Why it matters: 训练数据经常来自不同平台，字段名、对话角色和CSV列结构各不相同。这个工具提供了一个离线、本地、适合非技术用户的网页界面，把常见格式统一为： - `train.jsonl`：训练集 - `val.jsonl`：验证集 通过这个工具，用户可以轻松地整理和处理各种格式的训练数据，确保数据的

Overview: This library allows developers to convert text prompts into image representations, making it useful for various applications like image generation for text-based prompts. Key features: - Converts text prompts into image representations - Supports various text-to-image conversion - Provides a simple API for integration into existing Android projects Tech stack: Kotlin Why it matters: This library simplifies the process of creating image representations from text prompts, making it easier for developers to incorporate text-based image generation into their applications. Repository: zekierciyas/android-text-to-image GitHub stars: 11

Overview: GalaxDB is a database designed specifically for AI and ML workloads, providing a single binary with a single SQL interface for vector search, training, and training exports. Key features: - A single binary for AI-native database capabilities - One connection string, backup, and monitoring endpoint - Supports PostgreSQL wire protocol, making it compatible with existing SQL code - Integrates a relational database, vector store, embedding API, and training pipeline into one system - Supports semantic search, training datasets, and time-travel queries Tech stack: Rust Why it matters: GalaxDB simplifies the integration of AI and ML workloads by providing a unified database solution that eliminates the need for multiple services, making it easier to manage and scale applications. Repository: zentrix-innovative-labs/galaxdb GitHub stars: 11 GitHub description: GalaxDB is designed for AI and ML workloads that need more than a traditional database. Instead of stitching together a relational database, a vector store, an embedding API, and a training pipeline, GalaxDB provides all of these capabilities in one binary with a single SQL interface. README excerpt: <div align="center"> <img src="assets/Icon.svg" alt="GalaxDB" width="140" /> <h1>GalaxDB</h1> <p><strong>The AI-native database. SQL + vector search + training exports in one system.</strong></p> [![CI](https://github.com/zentrix-innovative-labs

### Overview: Strata v0.1.24 是一个用于在双 Tesla V100 sm_70 GPU 上运行的 Qwen3.8-Flash-Next 模型的官方补丁集合。目标是将 Strata v0.1.24（仅支持 RTX 30 系 sm_80+）在双 Tesla V100-16GB + 无 AVX-512 的老 Xeon 上正确运行，实测 decode 43.4 tok/s、prefill 865 tok/s @117K token 上下文。 ### Key features: - **补丁 01 (必需)**: 改善平台适配，支持双 Tesla V100 sm_70。 - **补丁 02 (必需)**: 修复正确性问题，防止输出乱码。 - **补丁 03 (可选)**: 优化双卡性能，提升 decode 速度。 - **补丁 04 (可选)**: 诊断开关，便于排查数值问题。 ### Tech stack: - Shell ### Why it matters: - 该补丁集合旨在让 Strata v0.1.24 在双 Tesla V100 sm_70 GPU 上正确运行，实测 decode 43.4 tok/s、prefill 865 tok/s @117K token 上下文。 - 通过打补丁 01、02、

Overview: Engram is an MCP server designed for AI memory, providing persistent personal memory with hybrid search, temporal decay, and local-first capabilities. Key features: - Persistent personal memory - Hybrid search (BM25 + semantic embeddings + knowledge graph) - Temporal decay modeled on the Ebbinghaus forgetting curve - Memory consolidation - Works with Claude, Claude Code, and any MCP client Tech stack: TypeScript Why it matters: Engram offers a solution for AI memory that goes beyond keyword matching, providing a more comprehensive and context-aware understanding of conversations and interactions. This technology is essential for creating a more effective and personalized AI experience.

### Overview: PotatoRAG is a hyper-optimized, privacy-first, local Retrieval-Augmented Generation (RAG) system designed to run on extremely low-resource hardware. It leverages Google's advanced TurboQuant algorithm and Turbovec vector index to compress high-dimensional text embeddings into ultra-lightweight 4-bit representations, ensuring lightning-fast local intelligence that fits in the palm of your hand. ### Key Features: - **TurboQuant Algorithm**: Utilizes Google's TurboQuant algorithm for ultra-fast vector indexing. - **Turbovec Vector Index**: Provides a lightweight 4-bit representation of text embeddings, reducing RAM usage by up to 80%. - **Local LLM Engine**: Runs `ollama` for generation and `nomic-embed-text` for embeddings, ensuring privacy and low resource usage. - **Streamlit Interface**: Offers a clean, interactive, and responsive web interface for user interaction. ### Tech Stack: - **Vector Database**: `turbovec` (SIMD-optimized local vector index based on Google's TurboQuant algorithm) - **Local LLM Engine**: `ollama` (Running `llama3.2:1b` for generation and `nomic-embed-text` for embeddings) - **User Interface**: `streamlit` (A clean, interactive, and responsive web interface) - **Data Operations**: `numpy` (For fast array manipulation and embeddings alignment) ### Why It Matters: - **Privacy**: Ensures no data leakage or cloud API usage. - **Resource Efficiency**: Runs on extremely low-resource hardware,

Overview: This repository contains a Jupyter notebook designed for fine-tuning language models using LoRA and 4-bit quantization techniques optimized for Kaggle's free tier GPU environment. The implementation focuses on memory-efficient training while maintaining model performance. Key features: - 4-bit Quantization - LoRA Fine-tuning - Kaggle Optimized - Memory Efficient - Clean Implementation Tech stack: Python 3.8+, CUDA-compatible GPU (T4 recommended) Why it matters: This script provides a practical solution for fine-tuning language models efficiently, making use of advanced techniques to optimize performance and memory usage, suitable for Kaggle environments. Repository: Abhijeet-ist/FineTunning GitHub stars: 10 GitHub description: This is a short script based on fine tuning a open sourced LLM based on customized parameters and data.

Overview: This project is a production-ready AI customer support chatbot built on Retrieval-Augmented Generation (RAG), advanced prompt engineering, and parameter-efficient fine-tuning. It answers customer questions accurately by grounding responses in your own documentation, ensuring policy-compliant answers. Key features: - RAG-powered chatbot - Advanced prompt engineering - Parameter-efficient fine-tuning (PEFT) Tech stack: Python Why it matters: This system ensures accurate and policy-compliant customer support, improving user satisfaction and reducing the risk of misinformation. Repository: AdilShamim8/Customer-Support-Chatbot-102 GitHub stars: 10 GitHub description: A fully production-ready AI customer support system built on Retrieval-Augmented Generation (RAG), advanced prompt engineering, hybrid search, reranking, and parameter-efficient fine-tuning.

Overview: This project is a lightweight, portable AI assistant designed for SAP/ABAP developers, offering features such as local AI integration, pre-built Windows setup scripts, a simple file-based knowledge store, and logging and outputs for debugging. Key features: - LocalAI integration for running models locally - Pre-built Windows setup scripts (PowerShell and batch files) - Simple file-based knowledge store for ABAP snippets and docs - Logging and outputs folder for debugging and traces Tech stack: Python 3.9+ Why it matters: This tool streamlines SAP/ABAP development by providing a local AI assistant, making it easier to learn, debug, and plan projects without the need for a cloud-based solution. Repository: Ankit-ku-panda/SAP-ABAP-LocalAI GitHub stars: 10 GitHub description: Portable offline AI assistant for SAP/ABAP learning, code drafting, debugging, and development planning. Runs locally with lightweight models, streaming responses, personal note retrieval, and optional GPU acceleration.

Overview: nano-vLLM is a small LLM inference engine designed to closely mimic the functionality of vLLM, focusing on understanding and replicating its core ideas. Key features: - An OpenAI-compatible `POST /v1/completions` endpoint that streams tokens over SSE. - Continuous batching of users joining and leaving between decode steps. - Our own SmolLM2 forward pass, including RMSNorm, RoPE, GQA, and SwiGLU. - A paged KV cache shared by all users, with a block table per user, avoiding re-padded or copied tokens. - Prefix caching to share blocks among users whose prompts start the same. - Chunked prefill for long prompts, reducing the time spent on waiting for tokens. Tech stack: Python Why it matters: nano-vLLM provides a detailed comparison of its performance metrics against a baseline server, showing significant improvements in throughput, decoding speed, and token efficiency. This makes it a valuable tool for developers looking to optimize their LLM implementations.

Overview: This project is an AI YouTube Video Chat application that allows users to ask questions to a bot powered by LangChain, a framework for language models. Key features: - Integration with LangChain for natural language processing - Use of MongoDB and Prisma ORM for database management - HNSWLIB vector store for efficient search capabilities - Next.js 13 app directory for complex state management and routing - NextAuth for Google authentication Tech stack: JavaScript, LangChain, MongoDB, Prisma ORM, HNSWLIB, Next.js 13 Why it matters: This application demonstrates the integration of cutting-edge AI technologies with traditional web development, showcasing how modern frameworks can be used to build sophisticated applications.

Overview: This project is an official PyTorch implementation of the paper "Where You Edit is What You Get: Text-Guided Image Editing with Region-Based Attention" by Pattern Recognition. Key features: - Text-Guided Image Editing - Region-Based Attention - PyTorch Framework Tech stack: Python Why it matters: This project offers a practical solution for text-guided image editing, leveraging region-based attention to enhance the quality of image modifications.

### Overview: MyLlama 是一个基于 llama.cpp 的跨平台本地大模型客户端，支持 Windows、Linux 和 Android 平台。它通过一键调优 GGUF 模型、共享 OpenAI 兼容端点、内置模型下载、本地聊天与实时监控等功能，为用户提供便捷的模型管理和高性能推理体验。 ### Key features: 1. **一键调优 GGUF 模型**：读取 GGUF 真实指标，结合硬件快照与实测内存带宽，自动规划推理参数。 2. **多模型共享 OpenAI 兼容端点**：一个 OpenAI 兼容端点供内置聊天与任意客户端使用，支持模型下载、本地聊天与实时监控。 3. **可视化配置与实时监控**：提供提示处理与生成 tok/s 的双指标，GPU、内存、CPU 实时采样，服务日志控制台按游标增量刷新、落盘不丢失。 4. **无头 API 模式**：一个开关切换为纯后台运行，提供健康检查、状态、日志与停止接口。 5. **一键智能调优**：读取 GGUF 真实指标，结合硬件快照与实测内存带宽，为每个模型自动规划推理参数。 6. **模型获取与管理**：支持 HF 镜像 + ModelScope 双源搜索，批量下载走可断点续传的任务队列；解析 GGUF 元数据，支持外部导入目录。 7.