Projects

### Overview: Diffusion-SDPO is a safeguarded direct preference optimization rule designed for diffusion models, ensuring that updates do not increase the winner's loss to first order while preserving the preferred output. ### Key features: - **Safeguarded Direct Preference Optimization (SDPO)**: Computes an adaptive scale for the loser branch based on the alignment between winner and loser output-space gradients, preventing the winner's loss from increasing. - **Model Agnostic**: Works with any diffusion model without additional training or modifications. - **Minimal Overhead**: Drops into Diffusion-DPO, DSPO, and DMPO with negligible performance impact. ### Tech stack: Python ### Why it matters: SDPO ensures that updates to diffusion models do not increase the loss of the winner, preserving the preferred output while maintaining a wide preference margin. This makes it a valuable tool for maintaining the quality and relevance of generated images in various applications.

Overview: This repository is a Multimodal Semantic RAG system designed to process documents with multimodal content, generating factual, context-aware answers based on textual, visual, and tabular information. Key features: - Semantic chunking - Image and table detection - Summarizing images and tables - Embedding - Semantic retrieval - Re-ranking - Answer generation Tech stack: Python Why it matters: This system aims to provide accurate, context-aware answers by integrating various data types, making it highly versatile and useful for applications that require handling diverse information sources. Repository: AhmedAl93/multimodal-semantic-RAG GitHub stars: 27 GitHub description: A RAG system designed to process documents with multimodal content. It can generate factual, context-aware answers to user queries, based on the documents texts, tables, figures, ...

Overview: Ear is a desktop app designed to transcribe audio from your computer, making it easier for those who cannot hear to understand what is being said. Key features: - Transcription capabilities - Desktop integration - Supports Windows, Linux, and Mac OS Tech stack: TypeScript Why it matters: Ear simplifies the process of transcribing audio, making it accessible to a broader audience, especially those who are deaf or hard of hearing. Repository: BrunoHenrique00/ear Primary language: TypeScript GitHub stars: 27

Overview: The Cognipeer Console is an open-source, self-hosted AI gateway and control plane designed for OpenAI-compatible LLM routing, RAG, an MCP hub, guardrails, and agent tracing for multi-tenant teams. Key features: - OpenAI-compatible LLM routing - Real-time WebSocket voice agents with barge-in and tool-call events - Model Hub & Dynamic Routing - MCP Hub for curating and hosting Model Context Protocol servers - Isolated runtime environments for code, agents, and previews - Managed browser profiles and live sessions - Agent-to-Agent publishing as A2A Protocol servers - Versioned templates for environment-based deployment - Document ingestion, chunking, embedding, and retrieval - Scheduled or ad-hoc web crawling - Multi-provider vector operations - Layout/table-aware document extraction and Cohere-compatible reranking - Scoped memory stores with vector-based recall Tech stack: TypeScript Why it matters: The Cognipeer Console is adopted by 15 companies as enterprise customers and is open-source for users to adopt and customize. It offers comprehensive features for AI gateway, agent automation, knowledge retrieval, observability, evaluation, safety, and infrastructure management, making it a valuable tool for multi-tenant teams.

Overview: Rummage 是一个基于 Google DeepMind EmbeddingGemma 2 的本地多模态文件语义搜索器，能够识别和搜索文档、照片、截图、录音和视频，并支持 Windows 和 Android 系统。 Key features: - 搜索文件内容，支持中文及其他100多种语言。 - 以图搜图，支持多种图像格式。 - 扫描相似文件，支持多种文件类型。 - 支持 Windows 和 Android 系统。 - 简洁的界面，包括主界面、搜索结果、设备检测、文件夹筛选等功能。 Tech stack: Rust Why it matters: Rummage 是一个高效、本地化的文件搜索工具，能够快速准确地找到与用户需求相关的文件，极大地提高了工作效率。它利用了 Google DeepMind EmbeddingGemma 2 的技术，实现了跨模态的搜索，支持多种格式的文件，提供了丰富的搜索功能，使得用户能够在不同的设备上轻松找到所需的信息。

### Overview: nano-lmcache is an educational project that demonstrates the principles of key-value cache reuse, prefix matching, and in-memory caching. It starts with a pure-Python prototype and progresses to real GPU tensors and integration with the nano-vLLM inference engine. ### Key Features: - **Prefix Matching**: Understand how to efficiently match keys. - **Chunk Keys**: Implement and use chunk keys for efficient storage. - **Storage Abstractions**: Explore in-memory caching before diving into more complex backends. - **GPU Integration**: Real GPU tensors verified against a real model, enabling GPU to CPU offload and a GPU -> CPU -> disk hierarchy. ### Tech Stack: Python 3.11+ ### Why It Matters: This project serves as a stepping stone for understanding key-value cache reuse, prefix matching, and caching principles. It builds upon the ideas presented in [LMCache](https://github.com/LMCache/LMCache), providing a clear path for learners to understand why certain design choices were made in production systems like LMCache and nano-vLLM. ### Repository: Dami-wq-k03/nano-lmcache ### GitHub Stars: 27 ### Primary Language: Python ### GitHub Description: An educational, step-by-step implementation of LLM KV cache reuse — learn prefix caching, chunk keys, storage & eviction, from pure Python to GPU tensors and vLLM integration. Inspired by LMCache.

### Overview: A curated collection of 100+ high-quality Nano Banana 2 prompts for AI image generation, including product photography, photorealism, street and documentary scenes, panoramic landscapes, 3D and dioramas, creative and artistic styles, and fashion and lifestyle elements. ### Key Features: - **Product Photography:** - **Orange Juice Brand Product Shot** - **Luxury Product Photography** - **Photorealism & Portraits:** - **Spherical Cute Animals** - **Needle-Felted Wool Miniature** - **Candid Street Portrait** - **Pixel Art Game Enemy** - **Hyper-Realistic Dual-Tone Portrait** - **Gritty Lion Bond Portrait** - **Soft Glow Mirror Selfie** - **Mosque Prayer Scene** - **Street & Documentary:** - **Black & White Urban Cyclist** - **White Horse in Motion** - **Minimalist Orange Wall** - **Blue Hour Child & Cat** - **Golden Hour Walking** - **Basketball Court Twilight** - **Vintage Indoor Portrait** - **iPhone Mirror Selfie** - **Panoramic & Landscape:** - **Mars Rover Panorama** - **Balcony River View** - **Jungle Cabin & Sunset Panorama** - **Enceladus Ice Surface Panorama** - **Snow Leopard at Dawn** - **3D & Diorama:** - **Hyper-Realistic City D

Overview: FortOS is a modern, security-first Linux NAS operating system built on .NET 10, designed for bare-metal deployment or Docker-based evaluation. It offers features such as disk discovery, storage protection, sharing capabilities, security features, and container management. Key features: - 🗄️ Storage: Supports ext4, XFS, Btrfs, and ZFS with per-share quotas. - 🔗 Sharing: Supports SMB, NFS, FTP, and a recycle bin with retention policies. - 🛡️ Protection: Provides snapshots, rsync backup, cloud backup, and point-in-time restore. - 🔐 Security: Offers NasToken capability tokens, fine-grained NAbility permissions, data classification, and HMAC-chained audit. - 🐳 Containers: Includes an Agent Catalog, Token Broker, and hardened Compose generation. - 📈 Observability: Offers metrics, a 5-stage log pipeline with Loki, alert engine, and Prometheus `/metrics` endpoints. Tech stack: C# Why it matters: FortOS is a robust and secure NAS operating system that combines modern Linux capabilities with advanced security features, making it ideal for businesses requiring high levels of data protection and management efficiency.

### Overview: iSDFT is an innovative approach to continual learning in large language models, specifically designed for information-proximal on-policy self-distillation. It leverages demonstrations to improve the student model's performance while maintaining a controlled drift towards the teacher's knowledge. ### Key Features: - **iSDFT Trainer**: The official implementation of iSDFT, including the trainer, datasets, and evaluation results. - **Tool Use Dataset**: Contains tool use tasks for demonstration-based learning. - **Science Dataset**: Features science-related tasks for knowledge transfer. - **Training Code**: The code for training the iSDFT model on both datasets. - **Evaluation Results**: Consistent with the paper, demonstrating improvements in various settings. ### Tech Stack: - Python 3.12 - CUDA GPU support ### Why It Matters: iSDFT addresses the challenge of continual learning in large language models by providing a method that balances demonstration learning with controlled drift towards teacher knowledge, leading to better performance across different benchmarks and tasks.

Overview: PatchFinder is a two-phase approach designed to trace security patches for disclosed vulnerabilities in open-source software. It leverages efficient initial retrieval via a hybrid retriever and powerful re-ranking using fine-tuning of CodeReviewer, significantly improving the recall and mean reciprocal rank of security patch tracing. Key features: - Efficient initial retrieval using a hybrid retriever - Powerful re-ranking using fine-tuning of CodeReviewer Tech stack: Python, TensorFlow, PyTorch Why it matters: PatchFinder addresses the challenge of efficiently tracing security patches for open-source software vulnerabilities, offering a more effective and scalable solution compared to traditional methods.

### Overview: `phones` is a Python library designed to handle International Phonetic Alphabet (IPA) phones, making it easier to describe and manipulate pronunciation in various languages. ### Key Features: - Extracts numeric feature vectors from phones. - Maps phones from one language to another by finding the closest phones. - Converts between ARPABET, X-SAMPA/SAMPA, and IPA notation. - Computes phone distances. - Performs phone arithmetic at both the phone and phone-feature level. - Provides visualization tools for phones and their distances when installed with `phones[plots]`. ### Tech Stack: Python ### Why It Matters: These IPA phones offer a standardized way to describe pronunciation, making it easier to communicate and compare phonetic representations across different languages and dialects.

### Overview: This repository is dedicated to AI/ML libraries and tools that are essential for developers and AI engineers. It includes popular Python libraries such as NumPy, Pandas, Scikit-Learn, XGBoost, LightGBM, TensorFlow, Keras, PyTorch, and Hugging Face's Transformers. These tools are crucial for building and deploying AI models and agents. ### Key Features: - **NumPy**: A fundamental library for numerical computations. - **Pandas**: A powerful data manipulation library. - **Scikit-Learn**: A popular machine learning library. - **XGBoost**: A fast, scalable, and high-performance gradient boosting library. - **LightGBM**: A fast and efficient gradient boosting library. - **TensorFlow**: A comprehensive library for machine learning and artificial intelligence. - **Keras**: A high-level neural networks API. - **PyTorch**: A popular deep learning library. - **Transformers (Hugging Face)**: A library for natural language processing. - **spaCy**: A library for natural language processing. ### Tech Stack: - Python ### Why It Matters: These libraries and tools are essential for developers and AI engineers who need to build, train, and deploy AI models and agents. They provide the necessary tools to handle numerical computations, data manipulation, machine learning, deep learning, and natural language processing tasks efficiently.

### Overview: Murmur is a local-first speech-to-text tool optimized for Chinese. It's more than "voice input" — press a hotkey to dictate and text appears at your cursor; import audio files for batch transcription and export; then let AI remove filler words or turn the transcript into meeting notes or a Xiaohongshu post. Everything happens on your computer, no internet, no need for an internet connection. ### Key Features: - **Local Processing**: Murmur runs entirely on your device, ensuring offline functionality. - **AI Polishing**: AI automatically cleans up the text, removing filler words and improving readability. - **Multi-File Support**: Import audio files for batch transcription and export. - **Hotkey Dictation**: Press a hotkey to dictate and text appears at your cursor. - **Customizable Hotkeys**: Set up custom hotkeys for different dictation styles. - **AI-Polished Transcripts**: Convert text into meeting notes or Xiaohongshu posts with AI polish. - **User-Defined Screenshot**: Customize the screenshot to your liking. ### Tech Stack: TypeScript ### Why It Matters: Murmur is a local-first speech-to-text tool optimized for Chinese, making it ideal for users who need to dictate text without an internet connection. The AI polishing feature ensures the text is clean and easy to read, making it perfect for various applications like meeting notes or social media posts.

### Overview: 本项目提供了一种方法，将一张图片逆向推导出生成提示词，包括中英文双语提示词、风格标签和结构化 JSON prompt。通过零依赖校验脚本，确保了输出的准确性与可靠性。 ### Key features: - **中英文双语重建提示词**：提供简体中文和英文版本的重建提示词。 - **风格标签**：输出中英文风格标签各4条。 - **结构化 JSON prompt**：提供12个字段的结构化提示词。 ### Tech stack: Python 3.8+ ### Why it matters: 本项目为图像分析和法证领域提供了一种创新的工具，通过零依赖校验脚本确保了输出的准确性与可靠性，适用于图像重建、法证分析和图像描述等场景。

Overview: This repository contains a collection of Jupyter Notebook files showcasing various machine learning techniques, data analysis, and data engineering tasks. Key features: - Anomaly Detection Using AutoEncoder with Pytorch - Bubble Detection for the Crypto Market - Data Engineering & Analysis - Depth Estimation 2D to 3D - Feature Importance with XGBoost vs ppscore - Fine-Tune LLaMA 2 with QLoRA for QA Datasets - Machine Learning Pipeline - McKinsey ProHack Challenge 2020 with XGBoost - NLP Cases - Quantize Llama 2.7B model and Fine-Tune on QA Datasets - Quantum Method for Portfolio Optimization - Rule-Based Turkish Language Sentiment Analysis - SARIMA Model - Turkey Conflicts Data Mining - Turkish Text Classification - Vector Databases Tutorial - Llama 2.13B Playground - Medical Trends Analysis - Portfolio Optimization - Stroke Classification - Tapu Kadastro Analysis Tech stack: Jupyter Notebook Why it matters: This repository provides a comprehensive set of machine learning projects and experiments, allowing developers to explore various techniques and gain practical experience in data analysis, machine learning, and data engineering.

Overview: LLM-KICK is a comprehensive evaluation protocol designed to assess the capabilities of compressed language models, focusing on their ability to understand, reason, generate text, and perform in-context retrieval and summarization tasks. Key features: - A collection of carefully curated tasks tailored to evaluate the performance of compressed LLMs. - It includes tasks that align closely with dense LLMs, emphasizing the importance of understanding and reasoning capabilities. - It provides insights into the limitations of current pruning methods and quantization techniques, highlighting the need for more effective strategies. Tech stack: Python Why it matters: LLM-KICK offers a standardized benchmark for evaluating the performance of compressed language models, ensuring that the evaluation metrics reflect the true capabilities of these models, especially in tasks requiring understanding and reasoning. This protocol is crucial for developers and researchers aiming to improve the performance and reliability of LLMs in various applications.

Overview: This project focuses on fine-tuning the black-box OpenAI embedding models to enhance their retrieval performance. Key features: - Fine-tuning the BAAI/bge-small-en model - Fine-tuning the FAE model Tech stack: Jupyter Notebook Why it matters: This method aims to improve the performance of black-box embedding models by fine-tuning them with additional information, thereby enhancing their utility and effectiveness in various applications. Repository: VectifyAI/Model-Augmented-Fine-Tuning Primary language: Jupyter Notebook GitHub stars: 27

Overview: TurboQuant Pro is a consumer-aware compression solution designed to optimize the performance and accuracy of embedding indexes and language model (LLM) key-value (KV) caches. It achieves this by compressing vectors based on the specific metric used by the downstream consumer, unlike traditional cosine similarity, which can be misleading. Key features: - **Consumer-Aware Compression**: Optimizes compression for specific consumer metrics like recall and attention quality. - **PCA-Matryoshka + TurboQuant**: Utilizes PCA-Matryoshka for key quantization and TurboQuant for embedding compression, achieving up to 27x compression with 99.8% recall@10. - **Asymmetric K/V**: Key quantization avoids the failure that is invisible to reconstruction metrics, improving key-value cache performance. - **At Scale & In Production**: Supports compressed-domain search, persisted sharded indexes, and drift monitoring. Tech stack: Python Why it matters: TurboQuant Pro addresses the limitations of traditional compression methods by focusing on the specific needs of the consumer, ensuring better performance and accuracy in real-world applications. Its consumer-aware compression and key quantization techniques have been validated through extensive benchmarks and real-world use cases, demonstrating significant improvements in recall and attention quality.

Overview: A Python voice AI assistant that listens to your microphone, transcribes speech to text using Google Gemini, and speaks the answers offline with pyttsx3. Key features: - Speech-to-text via Google Speech Recognition - LLM answers powered by Google Gemini - Text-to-speech replies with pyttsx3 - Exit phrases in Arabic / English - Configurable via .env file Tech stack: Python 3.11+ Why it matters: A focused demo of speech recognition, generative AI, and text-to-speech on Windows. Repository: ahmedbadr-dev/voice-ai-assistant Primary language: Python GitHub stars: 27 GitHub description: Python voice AI assistant: microphone STT, Google Gemini answers, offline TTS with pyttsx3. Clean env-based setup.

### Overview: KiroLink is a local proxy designed to intercept and replace standard CLI or BYOK authentication with Kiro/AWS credentials, allowing Claude Code to operate independently from Anthropic's native billing backend. ### Key Features: - **Local Proxy**: KiroLink acts as a local proxy to intercept and replace standard CLI or BYOK authentication. - **Kiro/AWS Credentials**: It uses Kiro and AWS credentials to authenticate and interact with the system. - **Independent Operation**: Claude Code can operate independently from Anthropic's billing backend, making it more flexible and secure. ### Tech Stack: - Go ### Why It Matters: KiroLink simplifies the integration of Claude Code with Anthropic's services, making it more accessible and secure for developers.