Projects

Overview: NanoVector is a lightweight, pure C99 implementation of vector search and episodic memory for AI agents, designed for embedded systems with minimal memory footprint. Key features: - Pure C99 - AVX2+FMA - ARM NEON - FASM x64 - Zero dependencies - ~120 KB in size Tech stack: Pure C99, AVX2+FMA, ARM NEON, FASM x64, Zero Dependencies Why it matters: NanoVector provides a highly efficient and compact solution for AI agents, making it ideal for embedded systems and small-footprint devices. Repository: eminsk/nanovector Primary language: Python GitHub stars: 7 GitHub description: The SQLite of Vector Search & Episodic Memory for AI Agents GitHub release: [eminsk/nanovector](https://github.com/eminsk/nanovector/releases) GitHub issues: [eminsk/nanovector](https://github.com/eminsk/nanovector/issues) GitHub discussions: [eminsk/nanovector](https://github.com/eminsk/nanovector/discussions) GitHub stargazers: [eminsk/nanovector](https://github.com/eminsk/nanovector/stargazers) GitHub issues: [eminsk/nanovector](https://github.com/eminsk/nanovector/issues) GitHub issues: [eminsk/nanovector](https://github.com/eminsk/nanovector/issues) GitHub issues: [eminsk/nanovector

### Overview: Mayaku is a real-time object detection, segmentation, and keypoint library designed for fine-tuning and exporting. It offers state-of-the-art performance across a variety of datasets, including high accuracy and efficient quantization. ### Key Features: - **Highest Fine-Tune AP at Nano Size:** Mayaku-n achieves an average precision (AP) of 58.9% across 100 datasets, surpassing RF-DETR nano (57.4%) and YOLO26n (54.8%). - **True Int8:** The deploy graph uses only convolutions, ReLU, add, max-pool, and nearest resize, allowing for end-to-end quantization. - **Runs and Deploys Everywhere:** No compiled extensions, deformable convolutions, or attention. Parity-tested exports to ONNX, Core ML, OpenVINO, and TensorRT. - **Detection and Instance Segmentation:** A single API and training loop for both detection and instance segmentation. ### Tech Stack: - Pure PyTorch - No custom CUDA kernels - Apache 2.0 license ### Why It Matters: Mayaku is designed to be the most accurate nano detector for fine-tuning on your own data. It provides high accuracy, efficient quantization, and compatibility across various hardware platforms, including true int8 on NPUs. The library is easy to install and can be used with just one `pip install` command. ### Repository: - **Repository:** emrahnazif/mayaku - **Primary Language

Overview: This project generates floorplan images from textual descriptions using a Conditional Generative Adversarial Network (GAN). Key features: - Custom dataset handling - Performance metrics: FID and IS - Configurable training options Tech stack: Jupyter Notebook Why it matters: This project offers a practical solution for generating floorplan images from textual descriptions, enhancing the efficiency and accuracy of architectural design.

Overview: This project is a self-implemented PyTorch implementation of some famous large language models (LLM) for inference purposes only. Key features: - KV caching - Sampling method (temperature, top_k, top_p, beam search) - Position Embedding (rotary position embeddings, partial rotary position embeddings) - Attention Mechanism (multi-head, multi-query, group-query) - Decoding (speculative decoding, Jacobi decoding, Look ahead decoding) - Context window extension (Positional Interpolation, NTK-Aware Scaled RoPE, ReRoPe, SelfExtend, YARN, StreamingLLM, Dynamic NTK) Tech stack: PyTorch Why it matters: This project serves as a practice and learning tool, not for serious applications. Repository: feixyz10/pytorch-llm-infer Primary language: Python GitHub stars: 7 GitHub description: pytorch llm inference: oh my implementation (from scratch) of some famous large language model (LLM) for inference purpose only.

Overview: This project implements a production-ready, scalable Retrieval Augmented Generation (RAG)-powered Large Language Model (LLM)-based context-aware Question-Answering (QA) application. Key features: - Vector similarity search within the embedding space using a vector database. - Relevant contexts passed to LLM. - LLM producing an answer aware of the relevant contexts. Tech stack: Python Why it matters: This project leverages advanced AI techniques to create a highly robust and scalable QA app, making it suitable for large-scale applications and providing a production-ready solution for context-aware question answering.

Overview: token.place is a peer-to-peer generative AI platform designed to match LLM users with volunteers who can donate spare compute resources. This allows for efficient and effective use of AI capabilities across the globe. Key features: - **Peer-to-Peer Matching**: Matches LLM users with volunteers for AI tasks. - **Secure Communication**: Ensures secure and reliable communication between users and volunteers. - **Easy Setup**: Quick and straightforward setup for both relay and full development environments. Tech stack: Python Why it matters: token.place offers a unique solution for leveraging AI resources efficiently, making it valuable for researchers, developers, and anyone looking to enhance their AI capabilities through peer-to-peer collaboration. Repository: futuroptimist/token.place GitHub stars: 7 GitHub description: Peer-to-peer generative-AI platform that matches LLM users with volunteers donating spare compute.

### Overview: FIndex is a project designed to explore and demonstrate practical applications of vector databases for face recognition and similarity search. It features two approaches: a service-oriented application utilizing Qdrant for efficient face similarity search, and a standalone console sample using local embedding generation and PostgreSQL with `pgvector` for a more straightforward demonstration. ### Key Features: - **Qdrant Service Application**: Uses `FIndex.Api`, `Findex.Embedder`, and Qdrant for face similarity search. - `FIndex.Api` exposes the public search endpoint. - `Findex.Embedder` generates embeddings through an HTTP `/embed` endpoint. - Qdrant stores and searches vectors. - `FIndex.Console` seeds Qdrant from local images and is manually launched as `seed-faces`. - **PostgreSQL / pgvector Sample**: Uses local embedding generation and PostgreSQL with `pgvector` for a standalone demonstration. - `FIndex.VectorDataSample` generates embeddings in-process using the ONNX model. - It stores and searches vectors in PostgreSQL with the `pgvector` extension. - It recreates its sample collection when run. - It is manually launched as `vector-data-sample`. ### Tech Stack: - **C#** (ASP.NET Core, ML.NET, ONNX Runtime, OpenCvSharp) - **Python** (FastAPI-based embedding service, InsightFace) - **Qdrant** (High-performance vector search engine) - **PostgreSQL** (Local PostgreSQL with `pgvector` extension) ### Why it Matters

### Overview: VectorAccelerate is a Swift library designed for high-performance GPU-accelerated vector operations, specifically optimized for Apple Silicon. Leveraging Metal 4 shaders, it provides up to 100x speedups for large-scale vector operations, making it ideal for applications requiring large-scale vector search and fine-grained control. ### Key features: - **Metal 4 Acceleration**: Unified command encoding and tensor operations for efficient GPU acceleration. - **Optimized Kernels**: Hand-tuned Metal shaders for specific dimensions (512, 768, 1536) for maximum performance. - **Memory Efficiency**: Advanced quantization and compression techniques for optimal memory usage. - **ML Integration**: Experimental learned distance metrics with MLTensor support for more complex ML pipelines. - **Seamless Integration**: Drop-in acceleration for VectorCore operations, ensuring compatibility and seamless integration. ### Tech stack: Swift6, Metal4, Metal4Context ### Why it matters: VectorAccelerate is crucial for applications requiring high-performance vector operations, especially in environments with limited storage, thermal, and power constraints. By leveraging Metal 4 and unified memory architecture, it delivers significant speedups, making it an essential tool for developers looking to optimize their vector-based machine learning applications.

### Overview: This repository provides a framework for fine-tuning the Coqui-AI's XTTS_V2 model, specifically tailored for multilingual text-to-speech applications. It offers tools for both standard fine-tuning and LoRA (Low-Rank Adaptation) based fine-tuning, enabling efficient and memory-efficient speech synthesis. ### Key features: 1. **Full Model Fine-tuning**: Fine-tuning the XTTS_V2 model using all available parameters. 2. **LoRA Fine-tuning**: Efficient fine-tuning using Low-Rank Adaptation for improved memory usage. 3. **Inference Scripts**: Tools to synthesize speech using both standard and LoRA-adapted models. 4. **Language Support**: Supports other languages supported by XTTS_V2. ### Tech stack: Python ### Why it matters: This framework simplifies the process of fine-tuning and adapting the XTTS_V2 model for various applications, ensuring efficient and effective text-to-speech synthesis.

Overview: This repository fine-tunes the `google/translategemma-4b-it` model for improving bidirectional English↔Welsh translations. Key features: - Fine-tuning of `google/translategemma-4b-it` for Welsh translations - Supports multiple hardware profiles (H200F, H200F2, H200F4, H200F8) - Uses a paper-aligned training configuration Tech stack: Python Why it matters: This fine-tuning improves the quality of English↔Welsh translations, making it useful for language translation and bridging the gap between different languages.

### Overview: The Universal NFT Vector Database is a project that aims to store a vectorized version of every image-based Ethereum NFT (ERC-721). This initiative leverages technology such as Alchemy API, Graph Protocol, and Pinecone to make it possible. The complete software infrastructure can be found in the repository, and it includes an analytics dashboard, search page, integration with the Graph Protocol and Alchemy API, backend server, RegNetY-080 image embedding, and a search API and visualization API. The project is designed for horizontal scalability and is scalable to handle hundreds of thousands of NFTs in a queue. ### Key features: - Vectorized storage of NFTs - Integration with Alchemy API and Graph Protocol - Scalable backend server - Image embedding for standardization - Search API for NFT similarity matching - Visualization API for image similarity visualization - Horizontal scalability for task queue ### Tech stack: Python 3.8+, Node.js v16+, MongoDB Community Edition ### Why it matters: This project is crucial for the development of a scalable and efficient NFT similarity matching system, making it easier for developers to find and analyze NFTs based on their images. The project's ability to handle a large number of NFTs in a queue and its integration with various APIs make it a valuable tool for the NFT community.

### Overview: 这个项目是一个基于Qwen2.5的佛经知识大模型LoRA微调项目，使用了CBETA电子佛典提取的经典原文，结合RAG增强和QwenAPI生成高质量问答数据。项目包括18部经典原文、12,285条高质量训练数据、专项数据增强功能，以及LoRA微调支持Jetson/GPU训练，旨在创建一个专业、如法、有边界且具有挑战性的佛学问答模型「善知识」。 ### Key features: - 18部经典原文 - 12,285条高质量训练数据 - 专项数据增强功能（幻觉控制、边界拒答、身份强化、唯识学专项） - LoRA微调支持Jetson/GPU训练 - 专业、如法、有边界 ### Tech stack: Python ### Why it matters: 这个项目为佛学研究和教育提供了高质量的训练数据和模型，有助于提升佛学知识的传播和理解。

Overview: This repository fine-tunes the LFM2-1.2B model for Korean-English bidirectional translation, achieving state-of-the-art performance by using GRPO+COMET and SFT training methods. The project is based on the LiquidAI model and has significantly outperformed 4B models. Key features: - Fine-tuning of LFM2-1.2B for Korean-English bidirectional translation - Utilization of GRPO+COMET and SFT training techniques - Achieving SOTA (State of the Art) performance on the benchmark Tech stack: Jupyter Notebook Why it matters: This project showcases the effectiveness of fine-tuning and advanced training methods in improving model performance, particularly in bidirectional translation tasks. Repository: gyunggyung/LFM2-KoEn-Tuning GitHub stars: 7 GitHub description: Fine-tuning LFM2-1.2B for Korean-English bidirectional translation. GRPO+COMET & SFT Training, outperforming 4B models.

Overview: AlignSQL is a project that aims to improve the accuracy of Natural Language to SQL (NL2SQL) translation by implementing a post-training pipeline that combines Zero-shot, Self-Consistency, and Deep Preference Optimization (DPO) techniques. This project is a continuation of the Qwen3-8B model's training and evaluation process, focusing on refining the model's ability to generate accurate SQL queries. Key features: - Zero-shot: Baseline performance on Spider dataset is 43.91% EX. - Self-Consistency: Achieves 74.3% EX by using majority voting on sampled candidates. - Deep Preference Optimization (DPO): Improves EX by 2.2% from 71.8% to 74.95%. - Candidate Oracle: Demonstrates that the bottleneck is in candidate selection rather than generation. Tech stack: The project is implemented using Python, with the Qwen3-8B model as the base, and utilizes the LLaMA-Factory framework for training and evaluation. Why it matters: This project is significant because it demonstrates the effectiveness of combining different post-training optimization techniques to improve the accuracy of NL2SQL translation, thereby enhancing the overall performance of the model.

### Overview: This n8n node allows integration of Fusionbrain.ai into workflows, providing high-resolution image generation capabilities. ### Key Features: - Text-to-Image: Generates images based on prompts and optional parameters. - List Models: Retrieves available image generation models. - List Styles: Customizes generated images with different styles. ### Tech Stack: TypeScript ### Why it Matters: This node is no longer maintained due to unstable API, suggesting alternative models for better performance and reliability.

### Overview: tabsearch is a Python package designed to solve the challenge of real-world mixed datasets, where traditional similarity search methods often fail. It provides a unified approach to search across text, categories, and numbers, making datasets searchable by meaning, category, and numbers. ### Key Features: - **Unified Search**: tabsearch can handle datasets with mixed types of data, including text, categories, and numbers, making it suitable for real-world applications. - **Automatic Type Detection**: It automatically detects the type of each column, ensuring accurate similarity calculations. - **Configurable Weights**: Users can adjust the focus of the search by emphasizing different types of data, such as text, numerical values, or categories. ### Tech Stack: - **Python**: The primary programming language used for development. ### Why It Matters: - **Real-World Usability**: tabsearch makes datasets searchable by meaning, category, and numbers, addressing the limitations of traditional similarity search methods. - **Efficiency**: It provides quick results, allowing users to see the results in just 60 seconds. - **Intelligent Automation**: Automatically detects the type of each column and configures weights for the search, making it more efficient and accurate. ### Repository: - **GitHub Repository**: [hariharaprabhu/tabsearch](https://github.com/hariharaprabhu/tabsearch) - **GitHub Stars**: 7 - **GitHub Description**: A package for Vector Search over a tabular dataset across multiple columns. ### Quick Start: ```bash pip install tabsearch ``` ### Example Usage:

### Overview: DaMoMe is a smart income and expense tracking app built with Kotlin Multiplatform, designed to simplify financial management and promote better money management habits. ### Key Features: 1. **Overview**: View your financial summary with daily, weekly, monthly, or all-time filters. 2. **EasyAdd**: Add transactions via manual input, image selection, or photo capture. 3. **Records**: View and search transaction history categorized by month and year. 4. **DaMommy**: A chatbot powered by Gemini API for personalized financial advice and insights. ### Tech Stack: Kotlin Multiplatform ### Why it Matters: DaMoMe aims to address the low financial literacy issue, offering an accessible solution to promote better money management habits. Its AI-powered chatbot and advanced financial advice can help users make smarter decisions and stay in control of their finances. Repository: [harissabil/DaMoMe](https://github.com/harissabil/DaMoMe) GitHub Stars: 7 GitHub Description: A smart income and expense tracking app built with Kotlin Multiplatform.

### Overview: This course provides a comprehensive introduction to modern Artificial Intelligence applications using Large Language Models (LLMs). Students learn fundamental concepts, practical applications, and advanced techniques for working with LLMs, including text classification, clustering, semantic search, prompt engineering, RAG systems, fine-tuning, and multimodal applications. ### Key Features: - **Text Classification and Clustering**: Implement text classification and clustering systems using pre-trained models. - **Prompt Engineering**: Design and optimize prompts for various NLP tasks. - **RAG Systems**: Build retrieval-augmented generation (RAG) systems. - **Fine-Tuning**: Fine-tune language models for specific tasks. - **Multimodal Applications**: Develop multimodal applications using vision-language models. - **Deployment**: Deploy and scale LLM applications in production environments. ### Tech Stack: Jupyter Notebook ### Why It Matters: This course equips students with the skills to understand and apply the fundamental concepts of LLMs, enabling them to build and deploy advanced AI applications. The practical, hands-on approach ensures that students can effectively implement and experiment with LLMs, making them valuable assets in the field of AI and machine learning.

Overview: This project is a basic "retrieval augmented generation" (RAG) application built with Langchain and OpenAI in Python, utilizing Pinecone for efficient similarity search and Gradio for the user interface. Key features: - Document processing - Embeddings generation with OpenAI - Efficient search with Pinecone - Integration with OpenAI language model - Question-Answer chain - User-friendly Gradio interface - Real-time results display - Extensibility and customization - Smart search inside PDFs Tech stack: Python Why it matters: This application leverages cutting-edge technologies to create an intelligent Q&A system, making it a valuable tool for tasks requiring natural language understanding and efficient document retrieval.

### Overview: Huzaifa Shoukat is a full-stack developer and AI expert known for building robust engineering solutions and cutting-edge AI applications. His projects include MCP servers, browser automation, and developer tools. He is skilled in Python and TypeScript, focusing on useful abstractions and safe automation. ### Key Features: - **x-use**: An MCP server for X research, outreach, and inbox management. No X API key required. - **Structured Inbox and Notifications**: Conversations, unread/read filters, message request folders, visible message context, and notification actors and links. - **Multiple Accounts**: Isolated browser sessions, per-account personas and proxies, action budgets, and recovery records for uncertain writes. - **Agent Integrations**: Claude Code plugin, skills for Claude Code and Codex, and stdio MCP support for Claude Desktop, Cursor, Windsurf, and other clients. - **Cross-Platform CI**: Windows, macOS, and Linux on Python 3.10–3.14, including installed-package checks and synthetic browser tests with Patchright and Playwright. ### Tech Stack: Python and TypeScript ### Why It Matters: Huzaifa Shoukat's projects are crucial for bridging the gap between robust engineering and cutting-edge AI, ensuring that the parts that matter after the demo end are well-defined and useful. His work on x-use and other projects demonstrates his ability to create practical, installable tools that developers can easily integrate into their workflows. ### Repository: GitHub Repository: [ihuzaifashoukat/ihuzaifashoukat](https://