AI Trending

Project directory

Projects

Open-source AI projects tracked by the leaderboard, ordered by GitHub stars.

Python 8 stars

doramirdor/mbolt

Profile-guided layout optimization for llama.cpp GGUF MoE models — BOLT/PGO for LLM weights. Trace expert routing, rewrite the file so co-activated experts sit together, prefetch as merged reads.

### Overview: Profile-guided layout optimization for MoE model files, reducing disk reads by 2.23x compared to the stock layout. ### Key Features: - **Rewrite Layout**: Reorders expert slices based on real routing traces to improve sequential access. - **Trace**: Logs which experts each token selects. - **Rewrite**: Reorders expert slices by measured co-activation. - **Rewrite**: Uses interleave layout for contiguous per expert blocks. - **Prefetch**: Explicitly reads selected slices to reduce page faults. ### Tech Stack: Python ### Why It Matters: Profile-guided optimization for MoE model files, reducing disk reads by 2.23x, improving inference performance.
Python 8 stars

duongtruongbinh/Math-Solving-Assistant

A chatbot website utilizing Gemini's API for solving math problems, chatting, and image analysis

Overview: Our team has developed a chatbot website called Math Solving Assistant, designed to assist users in solving mathematical problems, chatting, and analyzing images. The product is built using Gemini's API and Python, with Streamlit as the user interface. Deployed on the domain [mathsolving-chatbot.streamlit.app](https://mathsolving-chatbot.streamlit.app/), it features three main models: chat with Gemini, image analysis, and math problem assistance. Key features: - Chat with Gemini: Integrates Gemini's API to chat with users and convert image inputs into text for problem-solving. - Image Analysis: Analyzes uploaded images to describe their content. - Math Problem Assistance: Solves math problems using Gemini's model, with a fine-tuned model on a math dataset for improved accuracy. Tech stack: Python, Streamlit, Gemini API, Vertex AI. Why it matters: The product offers a comprehensive solution for mathematical problem-solving, image analysis, and enhanced chat functionality, making it a valuable tool for students and educators alike.
Jupyter Notebook 8 stars

faezeh-gholamrezaie/Fine-Tuning-Large-Language-Models-for-Sleep-Stage-Classification

Fine-tuning Large Language Models (LLMs) using QLoRA on EEG data for efficient sleep stage classification and pattern recognition.

### Overview: Fine-tuning Large Language Models (LLMs) using QLoRA on EEG data for efficient sleep stage classification and pattern recognition. ### Key features: - Fine-tuning large language models (LLMs) using QLoRA for sleep stage classification. - Utilizing QLoRA to efficiently fine-tune large pre-trained models like LLaMA. - Applying QLoRA to reduce memory footprint and computational requirements. - Optimizing input features from EEG data for effective classification. - Saving dataset in dictionary format for efficient storage and retrieval. ### Tech stack: Jupyter Notebook ### Why it matters: This project demonstrates the effectiveness of fine-tuning large language models on limited hardware resources, making it possible to classify sleep stages using EEG data without requiring high-end GPUs or extensive computing power.
Python 8 stars

firojalam/LlamaLens

This repository contains the resources, code, and documentation for LlamaLens, a specialized multilingual large language model (LLM) designed to analyze news and social media content effectively. LlamaLens supports multiple languages, including Arabic, English, and Hindi, and is tailored for diverse tasks such as sentiment analysis, misinformation.

### Overview LlamaLens is a specialized multilingual large language model designed for analyzing news and social media content. It supports Arabic, English, and Hindi, focusing on 18 NLP tasks across 52 datasets. ### Key Features - **Multilingual Support**: Arabic, English, and Hindi. - **Comprehensive Tasks**: 18 tasks utilizing 52 datasets. - **Domain Optimization**: Tailored for news and social media content analysis. ### Tech Stack - Python ### Why It Matters LlamaLens is optimized for diverse tasks such as sentiment analysis, misinformation detection, and more, making it a valuable tool for researchers and analysts working with multilingual news and social media data. ### Repository Repository: firojalam/LlamaLens GitHub Stars: 8 GitHub Description: Contains resources, code, and documentation for LlamaLens, a specialized multilingual LLM designed for analyzing news and social media content. ### Setup 1. Ensure Python and `pip` are installed. 2. Clone the repository. 3. Install required packages. 4. Customize the dataset preparation script using parameters such as `samples`, `split`, `intermediate_datasets_base`, `shuffling`, and `dataset_directory`. ### Features - **Multilingual Support**: Arabic, English, and Hindi. - **Comprehensive Tasks**: 18 tasks utilizing 52 datasets. - **Domain Optimization**: Tailored for news and social media content analysis. ### Dataset The model was trained on the LlamaLens dataset available on Hugging Face. ### Model
Jupyter Notebook 8 stars

geminii01/kaggle-genai-intensive-course

[Kaggle] 5-Day Gen AI Intensive Course with Google

### Overview: This Capstone Project is designed to showcase the culmination of the 5-Day Gen AI Intensive Course with Google, offering a practical application of the knowledge gained during the course. ### Key Features: - Apply and showcase what you've learned throughout the course - Submit a project that demonstrates your understanding and skills - Receive feedback and recognition from Kaggle and Google ### Tech Stack: Jupyter Notebook ### Why It Matters: This project provides a tangible way to demonstrate your progress and mastery of the course material, making it a valuable addition to your portfolio and a testament to your skills in Generative AI.
Unknown 8 stars

grknc/CNN-Guide-and-Resources

This repository provides an overview of Convolutional Neural Networks (CNNs) and how to get started with them using various datasets and techniques like transfer learning and fine-tuning.

Overview: This repository provides an overview of Convolutional Neural Networks (CNNs) and how to get started with them using various datasets and techniques like transfer learning and fine-tuning. Key features: - Introduction to CNNs - Explanation of the basic architecture - Detailed explanation of the Convolution Layer - Description of the Activation Layer - Explanation of the Pooling Layer - Explanation of the Flattening & Fully Connected Layers Tech stack: Python, TensorFlow, Keras Why it matters: CNNs are crucial for image and video recognition tasks, and this repository offers a comprehensive guide to understanding and implementing them effectively.
JavaScript 8 stars

halilxibrahim/ImageAI-App

Text To Image AI Mobile App

### Overview: This is a Text-to-Image AI Mobile App built with React Native, Expo, React Navigation, and Hugging Face.js. Users can input text and generate images using the AI model. ### Key features: - Login Screen - Home Screen - Text Input Screen - Image Generation - Image Display ### Tech stack: React Native, Expo, React Navigation, Hugging Face.js ### Why it matters: This app allows developers to quickly generate images from text, making it useful for various applications such as content creation, design tools, or even as a fun tool for personal projects.
Python 8 stars

haoruilee/DeepInfer-Mac

DeepSeek-V4.1-Flash (754B) on a single Apple Silicon Mac — DeepInfer-style MLX inference with SSD expert streaming, DSpark, SWA bounded replay, structured outputs

### Overview: DeepInfer-Mac is a project that re-imagines the DeepSeek model for inference on a single Apple Silicon Mac. It leverages SSD expert streaming, DSpark, and SWA bounded replay to achieve high performance and efficiency. ### Key Features: - **SSD Expert Streaming:** Efficiently reads and processes expert weights from a SSD. - **MLX Port:** Implements the official reference model on MLX, ensuring parity with the original. - **DSpark Verification:** Verifies the model's performance using speculative rejection sampling. - **Structured Outputs:** Allows for structured outputs using predefined schemas. - **Prefix Reuse:** Only the new tokens of a request feed into the decoder, avoiding redundant processing. - **Zero-Copy Reads:** Reads into the MLX slabs, reducing memory footprint. - **Confidence-Scheduled Verification:** Verifies the model's confidence levels for better performance. - **SWA Bounded Replay:** Only a bounded suffix of the prompt is replayed in subsequent layers. ### Tech Stack: - **Python:** The primary language used for development and inference. ### Why It Matters: This project aims to optimize the DeepSeek model for inference on Apple Silicon Macs, providing high performance and efficient use of memory and storage.
Python 8 stars

happycod3r/WinGPT

A cross-platform, fully functional, full-featured GUI implementation of the OpenAI API.

Overview: WinGPT is a fully functional GUI implementation of the OpenAI API, providing cross-platform compatibility and full control over API settings. Key features: - Cross-platform compatibility - Full-featured functionality including chat, file handling, audio transcription, image editing, image variations, translations, embeddings, fine-tuning, moderation, sentiment analysis, Q&A, summarization, and code generation - Control over all API settings through a modern GUI - Persistent and personalization settings for API key management and theme customization Tech stack: Python Why it matters: WinGPT simplifies the integration of the OpenAI API into a user-friendly GUI, making it accessible and customizable for developers and users alike.
Python 8 stars

harshilmathur/sparklab

DeepSeek Flash on 1-2 NVIDIA DGX Sparks, measured honestly: DeepSeek-V4.1-Flash-1017 on TensorFold TP2 (91.5 tok/s C1, 98% quality, 1M context), V4-Flash-0731 certified numbers, the laws, the harness, every CSV

### Overview: SparkLab is a benchmarking project measuring the performance of DeepSeek Flash models on NVIDIA DGX Spark hardware. This project has been meticulously documented over ten weeks, including raw CSV data, engine tags, and multiple repeats for accuracy. ### Key Features: - DeepSeek Flash model performance on two DGX Spark boxes. - Measured on TensorFold TP2 with EXL3 2.9 bpw weights. - Daily driver: DeepSeek-V4.1-Flash-1017, achieving 91.5 tok/s on single stream. - Metrics include single-stream decode, aggregate, time to first token, quality, and cold prefill performance. - Stability and load time are also recorded, with a 24-hour soak test showing no failures. ### Tech Stack: Python ### Why it Matters: This repository serves as a measurement layer for DeepSeek Flash models, providing benchmark data that can be used to validate and improve recipes for agent-shaped workloads. The raw, unaltered numbers from this project are crucial for understanding the performance of DeepSeek models on real hardware, which is essential for developers and researchers alike.
Python 8 stars

hedayat-atefi/rag-chroma-langchain

### Overview: This repository provides a developer-focused example and template for integrating Retrieval-Augmented Generation (RAG) with LangChain, Chroma vector store, and embeddings. It includes a small LangChain template package and examples demonstrating how to ingest documents, create a Chroma collection, and build a RAG pipeline. ### Key Features: - **Document Loading**: Supports URL and file uploads. - **Document Chunking and Text Splitting**: Ensures text is split appropriately for processing. - **Vectorstore Creation with Chroma**: Efficiently stores and retrieves documents using Chroma. - **Embedding Documents**: Provides a provider-agnostic approach to embedding documents. - **RAG Chain Scaffolding**: Facilitates context-aware question answering using LangChain. - **Reusable LangChain Template Package**: Simplifies the integration of LangChain projects. ### Tech Stack: - Python 3.8+ - Pip - LangChain - LangChain Core - Chroma vector store (via chromadb) - Embedding providers (OpenAI, Cohere, etc.) ### Why It Matters: This template is designed to be a comprehensive guide for developers looking to implement RAG using LangChain and Chroma, providing a solid foundation for integrating these powerful tools into their projects. ### Repository Layout: - **README.md**: This file. - **packages/rag-chroma**: A reusable LangChain template package with chain, ingest utilities, and a small FastAPI-based example. - **scripts/**: Small CLI utilities for ingesting files and URLs into a Chroma collection. - **note
Unknown 8 stars

hofong428/Optimizing-GPU-Kernels

LLM Serving & Inference Optimization

### Overview: Optimizing GPU Kernels is crucial for enhancing the performance of deep learning models and scientific computing applications. This project aims to explore and optimize GPU kernels using Triton, a programming language and compiler designed to simplify GPU kernel development. ### Key Features: - **Triton**: A new programming language and compiler for GPU kernels, offering higher-level abstractions and reduced complexity. - **CUDA**: A mature and widely used GPU programming language with extensive libraries and community support. - **Performance**: Preliminary tests indicate that Triton can match or exceed the performance of hand-written CUDA code for certain tasks, especially in ease of use and development speed. ### Tech Stack: - **Triton**: A programming language and compiler for GPU kernels. - **CUDA**: A mature and widely used GPU programming language with extensive libraries and community support. ### Why It Matters: Triton simplifies the development of GPU kernels, making it easier for researchers and developers to implement and optimize their code without the complexity of CUDA. This project aims to demonstrate the potential of Triton in improving the performance and development speed of GPU kernels, particularly for deep learning training and inference tasks. ### Repository: hofong428/Optimizing-GPU-Kernels ### GitHub Stars: 8 ### GitHub Description: LLM Serving & Inference Optimization
Rust 8 stars

huahuadeliaoliao/DeepSeek-OCR-2-burn

A Rust implementation of DeepSeek-OCR-2 using Burn. | 基于 Burn 框架的 DeepSeek-OCR-2 Rust 实现

Overview: This repository is a Rust implementation of DeepSeek-OCR-2 using the Burn framework, focusing on loading Hugging Face weights and end-to-end OCR inference. It aims to provide a faster and more memory-efficient solution compared to mature Python stacks. Key features: - Loading Hugging Face weights into Burn - End-to-end OCR inference in Rust - Support for both Vulkan (wgpu) and NdArray (CPU) backends Tech stack: Rust Why it matters: This implementation aims to offer a faster and more memory-efficient solution compared to mature Python stacks, particularly in scenarios where GPU acceleration is desired.
TypeScript 8 stars

ikhsan3adi/RuwetMeter

AI Agent buat mantau keruwetan Indonesia. Scrape berita, analisis pakai LLM. Bertanya ke chatbot RAG interaktif.

Overview: RuwetMeter is an AI-driven system designed to monitor and analyze socio-political stability in Indonesia, providing real-time insights and a chatbot for interaction. Key features: - Continuous scraping of news from major Indonesian portals - Sentiment analysis across 4 dimensions (economy, politics, infrastructure, social) - Real-time dashboard with anomaly detection and a RAG-powered chatbot - Automatic aggregation of RSS feeds and article extraction - Integration with PostgreSQL for metrics and vector search - Customizable visualizations including bar charts and trend lines - Anomaly detection based on delta changes Tech stack: TypeScript Why it matters: This project offers a comprehensive tool for monitoring socio-political stability in Indonesia, providing valuable insights and a user-friendly interface for interaction. Repository: ikhsan3adi/RuwetMeter Primary language: TypeScript GitHub stars: 8 GitHub description: AI Agent buat mantau keruwetan Indonesia. Scrape berita, analisis pakai LLM. Bertanya ke chatbot RAG interaktif.
Swift 8 stars

impel-intelligence/MagnitudeDB

A fast and modern Swift Vector Database for use in both iOS and macOS applications.

Overview: MagnitudeDB is a fast and modern Swift Vector Database optimized for iOS and macOS applications, backed by an SQL database and FAISS for efficient searching, with configurable storage locations and a collection system for organizing document groups. Key features: - Backed by an SQL database - Backed by FAISS for fast and efficient searching - Configurable storage locations - Collection system for organizing document groups Tech stack: Swift Why it matters: MagnitudeDB simplifies vector database development by providing a fast, modern solution for iOS and macOS applications, enhancing performance and organization through SQL and FAISS integration.
JavaScript 8 stars

inboxpraveen/Easy-vLLM

Choose a model and GPU, estimate memory use, and generate a vLLM command and Docker setup. Runs in your browser.

### Overview: Easy-vLLM is a user-friendly tool for quickly setting up and running a vLLM server. It allows users to choose a model and GPU, estimate memory usage, and generate a vLLM command and Docker setup without any installation or account required. ### Key features: - **Model Selection**: Choose any model from Hugging Face, including Mixture-of-Experts models. - **GPU Configuration**: Select the GPU type and number of GPUs available. - **Memory Estimation**: Estimate how much memory a model will use on the chosen GPU. - **Command Generation**: Automatically generate a vLLM command and Docker setup. - **Browser-Only**: Runs entirely in the browser, with no need to install anything. ### Tech stack: JavaScript ### Why it matters: Easy-vLLM simplifies the process of setting up a vLLM server, making it accessible to users without any technical background. It answers the three most common questions about running a model on a GPU, ensuring that users can quickly start using their models without needing to understand complex flags or configurations.
Python 8 stars

infino-ai/llama-index-vector-stores-infino

LlamaIndex integration for Infino — vector, BM25, hybrid (RRF), and SQL retrieval over one engine on object storage.

### Overview: Infino is a vector, full-text (BM25), hybrid (RRF), and SQL-native retrieval engine integrated with LlamaIndex, designed for efficient querying over object storage. This package allows seamless integration of LlamaIndex with Infino, providing a unified vector store surface for all retrieval modes. ### Key Features: - **Unified Vector Store Surface**: All four retrieval modes (vector, pure BM25, RRF hybrid, MMR) are exposed as first-class paths on the same table. - **Single Engine**: Runs SQL, BM25, vector, and hybrid (RRF) retrieval over a single in-process engine, eliminating the need for separate search clusters or vector stores. - **Durable Storage**: Data is kept in Apache Parquet on object storage, ensuring data durability and scalability. ### Tech Stack: - Python 3.10+ - Dependencies: `infino`, `llama-index-core`, `pyarrow`, `numpy` ### Why It Matters: Infino leverages the power of LlamaIndex while providing a more comprehensive and efficient querying experience. By integrating LlamaIndex with Infino, users can leverage the vector store surface of LlamaIndex while benefiting from Infino’s object storage capabilities, including full-text, hybrid, and SQL-native retrieval. This package simplifies the integration process, making it easier to build and deploy vector databases on object storage. ### Repository: Repository: infino-ai/llama-index-vector-stores-infino GitHub Stars: 8 GitHub Description: LlamaIndex integration for Inf
Python 8 stars

intel/document-level-sentiment-analysis

Document Level Sentiment Analysis is an End-to-End deep learning workflow using Hugging Face transformers API to do a "classification" task at document level, to analyze the sentiment of input document containing English sentences or paragraphs.

Overview: Document Level Sentiment Analysis is an end-to-end deep learning workflow using Hugging Face transformers API for analyzing the sentiment of input documents containing English sentences or paragraphs. This project is designed to classify sentiments at the document level. Key features: - Fine-tuning and inference pipeline for document-level sentiment analysis - Supports single node and multinode implementations - Utilizes BERT model from Hugging Face transformers API - Optimized for Intel hardware Tech stack: Python Why it matters: This project is a valuable tool for sentiment analysis in large document sets, making it useful for applications such as customer feedback analysis, market research, and more. Repository: intel/document-level-sentiment-analysis Primary language: Python GitHub stars: 8
Python 8 stars

intel/enterprise-inference

Enterprise-grade, high-performance LLM serving on Kubernetes. OpenAI-compatible inference powered by vLLM and OpenVINO™ Model Server behind a unified AI gateway, with NUMA-aware CPU optimization for Intel® Xeon® - production model serving, routing, and security built in.

### Overview: Introducing Intel® Inference Toolkit, a comprehensive solution for deploying and managing enterprise-grade LLMs on Kubernetes. This toolkit simplifies the process of serving models, providing a unified AI gateway, NUMA-aware CPU optimization, and OpenAI compatibility. ### Key Features: - **Unified AI Gateway**: Simplifies model routing and authentication. - **NUMA-Aware CPU Optimization**: Ensures optimal performance on Intel® Xeon® CPUs. - **OpenAI Compatibility**: Provides an OpenAI-compatible endpoint for seamless integration. - **Model Management**: Manages model versions and configurations. - **CLI for Model Management**: `model-manager` CLI simplifies the transition from Hugging Face to OpenAI-compatible endpoints. ### Tech Stack: - Python - Kubernetes - vLLM - OpenVINO™ Model Server - NumPy - OpenCV - TensorFlow - PyTorch - Transformers - Hugging Face ### Why It Matters: - Simplifies the deployment of enterprise-grade LLMs on Kubernetes. - Provides a unified and optimized model serving environment. - Ensures high performance and compatibility with OpenAI standards. - Streamlines the management of model versions and configurations. ### Repository: Intel/enterprise-inference ### GitHub Stars: 8
Shell 8 stars

irresi/awesome-agentic-knowledge-base

Empirical map of 46 trending AI agent repos: storage, memory, MCP, retrieval — what they actually use, ranked by adoption

### Overview: This repository is an empirical map of 46 trending AI agent repos, focusing on storage, memory, MCP, and retrieval. It ranks repos by adoption frequency, not vendor claims. ### Key features: - 21 of 46 (45%) use LLM-based entity extraction, but 4 ship serviceable KBs without any LLM cost. - 18 of 46 expose MCP servers, 17 are clients, and 10 intentionally avoid MCP. - Postgres leads metadata, pgvector leads vector backends, and Redis is the default cache. - 12 repos use no graph at all, with in-process NetworkX being more common than Neo4j. - License hardens with deployment shape: library → product → infra mirrors MIT → AGPL → ELv2/SSPL. ### Tech stack: Shell ### Why it matters: This repository provides a comprehensive overview of the adoption patterns of AI agent repos, offering insights into the actual technologies and practices used by developers.