Projects

### Overview: This repository provides the code for a multimodal realistic simulation framework using CARLA and MATLAB. CARLA is used to generate realistic autonomous driving sensor data, while MATLAB is employed for wireless communication simulations. This framework enables multimodal sensing data collection, 3D digital reconstruction, and ray-tracing-based wireless channel modeling to facilitate beamforming and channel analysis. ### Key Features: - **Realistic multimodal sensing**: LiDAR, Radar, and RGB Cameras integrated into an urban driving environment. - **Seamless 3D environment conversion**: CARLA to MATLAB using Blender API. - **Accurate wireless channel modeling**: Ray tracing for beamforming evaluation. ### Tech Stack: Python ### Why It Matters: This framework facilitates advanced beamforming and channel analysis in complex urban environments, enabling more effective sensing and communication in autonomous driving scenarios.

### Overview: MAMKit is a comprehensive toolkit designed for multimodal argument mining, focusing on the classification of arguments in speech transcripts and audio recordings. It provides a unified framework for running experiments on various corpora, configurations, and metrics, making it highly reproducible and easy to use. ### Key Features: - **Unified Configuration System**: All experiments are registered tasks, allowing for easy reproduction and comparison. - **Corpora and Featurizers**: Components like corpora, split strategies, featurizers, models, and metrics are registered with specific configurations. - **Model Training and Evaluation**: Supports training and evaluation of models on various input modes and datasets. ### Tech Stack: - **Python**: The primary programming language used for developing and executing tasks. - **Cinnamon**: The underlying toolkit on which MAMKit is built, providing a modular and extensible architecture. ### Why It Matters: MAMKit simplifies the process of multimodal argument mining by providing a standardized and reproducible environment. It ensures that experiments are easily replicable and comparable, making it a valuable tool for researchers and practitioners in the field of argument mining. ### Repository: The repository is hosted on GitHub at [nlp-unibo/mamkit](https://github.com/nlp-unibo/mamkit). It has received 14 GitHub stars, indicating its popularity and usefulness in the community. ### Installation: - **Python Version**: Requires Python 3.10 or later. - **Dependencies**: FFmpeg is required for audio tasks. - **Command Line Interface**: The `cmn-run` command

Overview: Nodaro is an API-first studio designed for AI media workflows, enabling the chaining of 100+ models including Seedance, Veo, Kling, Flux, ElevenLabs, and Suno, into node-based pipelines that can be run on schedules, webhooks, and Telegram, and published to 20+ platforms. Key features: - 100+ node types across 16 categories - 100+ AI models behind one interface - Workflows are portable JSON - Server-side execution engine - Workflows as products - Entity studios - Programmatic rendering - Self-hostable in two commands Tech stack: TypeScript Why it matters: Nodaro simplifies the complex process of AI media production into a user-friendly interface, allowing users to create, share, and execute workflows without the need for complex APIs or cloud accounts. This makes it accessible and efficient for both developers and non-technical users alike.

Overview: This repository offers a series of beginner-friendly projects aimed at teaching the fundamentals of Generative AI, including Prompt Engineering, LLM APIs, Embeddings, Vector Databases, RAG, and AI Agents. It provides practical, hands-on learning through real-world projects. Key features: - Prompt Engineering - LLM APIs (e.g., OpenAI API) - Embeddings - Vector Databases - RAG (Retrieval-Augmented Generation) - AI Agents Tech stack: Python 3.10+ Why it matters: This repository is designed to help beginners understand and build practical projects related to Generative AI, making it easier to transition from a beginner to an intermediate level in the field.

Overview: Vectrain is a high-performance, modular, plug-and-play RAG pipeline built in Go, designed for efficient ingestion, embedding, and storage of vector embeddings for semantic search, recommendations, and analytics. Key features: - Modular system with source, embedder, and storage components - Supports Kafka and REST data sources - Generates vector embeddings using Ollama models - Stores embeddings in Qdrant database - Configurable batch sizes and worker counts - Provides RESTful API for controlling and interacting with the pipeline Tech stack: Go Why it matters: Vectrain simplifies the integration of vector embeddings into various applications, enhancing semantic search, recommendations, and analytics capabilities with its high-performance, scalable architecture.

Overview: This project is a demonstration of how to diagnose and apply different retrieval methods for various purposes using Pinecone, focusing on improving RAG (Retrieval-Augmented Generation) retrieval. Key features: - A Marimo notebook that teaches when to use different retrieval techniques, including keyword, dense semantic, visual cross-modal, and combining them. - Real Pinecone calls for each retrieval method. - A small hand-built evaluation set. Tech stack: Python Why it matters: This project helps developers understand when to apply different retrieval methods for various purposes, enhancing their ability to diagnose and apply retrieval techniques effectively. Repository: pinecone-io/better-rag-retrieval-marimo Primary language: Python GitHub stars: 14 GitHub description: Learn how to diagnose and apply different retrieval methods for different purposes with Pinecone.

### Overview: Lo, a tiny code tutor that runs your code instead of guessing ### Key features: - Fine-tune Qwen2.5-Coder-7B for free and run a local coding agent that *runs your code instead of guessing*. - 8 GB Mac, llama.cpp, LoRA weight analysis with charts. - 20 GitHub stars, Python primary language, free training data, verified results. ### Tech stack: Python ### Why it matters: - Every student deserves a private, free code tutor that explains why their code behaves the way it does. - A 7B model on an 8 GB laptop can provide a patient explanation of language mechanisms, why step by step, and fix the trap. - Runs whole programs, data analysis, and looks inside neural networks with charts. - Reproducible for $0, a recipe anyone can repeat: free GPU for training, local inference, every decision and measurement documented. ### Repository: pnnnx/local-code-tutor ### GitHub stars: 20 ### GitHub description: Fine-tune Qwen2.5-Coder-7B for free (QLoRA on Kaggle T4) and run a local coding agent that runs your code instead of guessing — 8 GB Mac, llama.cpp, LoRA weight analysis with charts

### Overview: A comprehensive, mission-critical AI-powered emergency response ecosystem designed to revolutionize crisis management across the entire emergency response chain. From citizen reporting to hospital coordination, this platform delivers intelligent, offline-first emergency response tailored for high-stakes, low-connectivity environments where every second counts. ### Key features: - **Gemma 3n Integration** – State-of-the-art emergency analysis with 95%+ accuracy. - **128K Context Window** – Complete situational awareness across multiple incidents. - **Multi-modal AI** – Simultaneous text, voice, image, and sensor data processing. - **Edge AI Processing** – Critical analysis runs locally when connectivity fails. - **Continual Learning** – System improves with each emergency response. - **Voice Command System** – Hands-free operation with 95% accuracy across 15+ languages. - **Conversational AI** – Natural language queries for complex data analysis. - **Emergency Voice Detection** – Automatic stress and urgency level identification. - **Multi-speaker Recognition** – Distinguishes between different team members in chaotic environments. - **Real-Time Intelligence Engine** – Live patient assessment, predictive analytics, emergency operations intelligence, risk stratification, and pattern recognition. - **Offline-First Design** – Full functionality without internet connection. - **Mesh Networking** – Device-to-device communication when infrastructure fails. - **Satellite Integration** – Backup connectivity via Starlink and emergency satellite networks. - **5G/LTE Optimization** – Intelligent bandwidth management for critical data priority. - **Edge Computing

### Overview: Fine-tune a local Large Language Model (LLM) on your own app's data in just 15 minutes. Runs entirely on-device, zero cloud after training. Compatible with Apple Silicon and NVIDIA GPUs. ### Key features: - **Local Fine-Tuning**: Fine-tune a 4B model locally with LoRA. - **Zero Cloud After Training**: No need to upload your data to the cloud. - **RAG Pipeline**: Automatically generate training data using your app's existing RAG pipeline. - **Claude Annotation**: Use Claude to annotate gold-standard responses. - **On-Device Serving**: Serve the fine-tuned model on-device, forever, for free. ### Tech stack: Python ### Why it matters: This cookbook demonstrates a practical approach to fine-tuning an LLM locally, leveraging your app's existing data and infrastructure. It eliminates the need for cloud-based services, making it ideal for developers looking to optimize their app's performance and reduce costs.

Overview: This repository provides a Python script for fine-tuning the open-source Video Vision Transformer (ViVit) model using the HuggingFace Trainer Library. The model is pre-configured for 10 classes and can be used to classify videos into specific categories. Key features: - Fine-tuning of ViVit model - Pre-configured for 10 classes - Python script implementation Tech stack: Python Why it matters: This script is useful for researchers and developers looking to fine-tune ViVit for specific video classification tasks, making it easier to adapt the model to their needs.

Overview: This repository features 14 GitHub stars and showcases real-world AI projects using machine learning techniques. Key features: - Implementing image classification using Convolutional Neural Networks - Developing a speech recognition system using Natural Language Processing - Creating a recommendation system using collaborative filtering - Building a fraud detection model using anomaly detection Tech stack: Jupyter Notebook Why it matters: These projects demonstrate the practical application of machine learning in addressing real-world challenges, making them valuable for developers looking to integrate AI into their projects.

### Overview: Chimera is a single-executable library that combines llama.cpp, whisper.cpp, and stable-diffusion.cpp into a unified binary. It provides a single ggml-backed model backend, shared SQLite database, and OpenAI-compatible HTTP server across text, audio, and image modalities. ### Key Features: - **Unified Backend**: All engines share a single ggml backend for efficient inference. - **Shared Persistence**: Chat history, embeddings, and vector-store collections are stored in a single SQLite database. - **OpenAI-Compatible HTTP**: Provides routes for chat, embeddings, audio transcription, image generation, and RAG. - **Static Library**: Produces `libchimera.a` for embedding into other C++ processes. ### Tech Stack: C++ ### Why It Matters: Chimera simplifies the process of running multiple ggml-backed models in a single process, sharing a single ggml backend, SQLite database, and OpenAI-compatible HTTP server. It is particularly useful for users who want faithful upstream flag coverage, distribute a single static binary, and build on top of an OpenAI-compatible HTTP server. ### GitHub Stars: 14 ### Repository: shakfu/chimera

### Overview: MGFlow is an official implementation of the "Unifying Distributional Training for One-Step Visual Generation" paper. It post-trains one-step visual generators by matching real and generated features in frozen representation spaces. The framework connects distributional objectives to per-sample updates through Wasserstein gradient flow, using Gaussian mixtures for componentwise statistics and mass-constrained assignments to keep generated and reference components aligned. ### Key features: - **Distributional Training**: MGFlow post-trains one-step visual generators by matching real and generated features in frozen representation spaces. - **Gaussian Mixtures**: Provides componentwise statistics for Wasserstein or KL matching. - **Mass-Constrained Assignments**: Keeps generated and reference components aligned. - **Supported Tasks**: ImageNet 256×256, Text-to-image (FLUX.2 4B), and evaluation assets. ### Tech stack: Python ### Why it matters: MGFlow offers a unified approach to distributional training, enabling efficient generation of high-quality images and text-to-image samples. Its ability to match real and generated features in a single step makes it particularly effective for applications requiring precise control over the generated output. ### GitHub stars: 13 ### Repository: [shihaoyang0423/MGFlow](https://github.com/shihaoyang0423/MGFlow)

Overview: This serverless application utilizes Stable Diffusion XL to perform a Text-to-Image task on RunPod, offering a cost-effective solution for rapid responses without the need for idle time. Key features: - Text-to-Image functionality - Supports custom models - Compatible with low VRAM GPUs - Detailed instructions for deployment and usage Tech stack: Python Why it matters: This application is ideal for applications that require quick responses without the need for idle time, making it suitable for frequent but low-usage tasks. Repository: sky24h/SDXL_Serverless_Runpod GitHub stars: 14

### Overview: "Make-A-Video" is a cutting-edge project that leverages the power of Meta-FAIR's Tensorflow to generate videos directly from text descriptions without requiring any text-to-video data. This innovative approach aims to democratize video creation by enabling anyone to produce short, engaging videos quickly and easily. ### Key Features: - **Text-to-Video Generation**: Converts textual descriptions into video sequences. - **Spatio-Temporal Convolution**: Extends the model's building blocks to handle temporal information. - **Frame Interpolation Network**: Generates high frame rates for smoother video playback. ### Tech Stack: Python ### Why It Matters: This project represents a significant advancement in the field of video generation, offering a practical solution for those seeking to create videos without the need for extensive training data or specialized equipment. It has the potential to revolutionize the way videos are created and consumed, making it accessible to a broader audience.

### Overview: ReadmeReady is a tool designed to automatically generate README documentation for code repositories, addressing the challenges of manual documentation and the financial costs associated with using OpenAI APIs. ### Key Features: - **Auto-Generated README**: Quickly create basic documentation for any publicly available or custom repository. - **Fine-Tuned Models**: Utilizes a large language model (LLM) to generate READMEs, offering a fine-tuned model and features for custom tuning. - **Customizable**: Allows users to set custom parameters such as audience, link hosting, and more. - **Easy Installation**: Available on Linux/Windows, with dependencies installed via pip or cloning from source. - **Virtual Environment**: Recommended to create a virtual environment before installing. ### Tech Stack: Python ### Why it Matters: ReadmeReady simplifies the process of generating documentation, making it accessible and cost-effective for developers. It addresses the challenges of traditional manual documentation and the financial burden of using OpenAI APIs, offering a practical solution for modern software development practices.

Overview: A Python application designed to scrape news articles, generate summaries using GenAI, and perform semantic search through vector embeddings. Key features: - Scrapes news articles - Generates summaries - Identifies topics - Provides semantic search capabilities Tech stack: Python Why it matters: This application combines AI news scraping with semantic search, offering a comprehensive solution for news analysis and information retrieval.

Overview: The SUMS dataset, collected by Qinghai University, is a comprehensive dataset of 80 synchronized non-contact facial and contact finger videos from 10 subjects during exercise and oxygen recovery scenarios. It includes PPG, respiration rate (RR), and SpO2 measurements, designed to validate video vitals estimation algorithms and compare facial PPG with finger CPG. Key features: - Comprehensive dataset of 80 synchronized videos - Includes PPG, RR, and SpO2 measurements - Designed for validating video vitals estimation algorithms and comparing facial PPG with finger CPG Tech stack: The dataset is primarily in Python, utilizing specific libraries and tools for video processing and analysis. Why it matters: The results demonstrate significant improvements in SpO2 prediction accuracy by fusing videos from different positions (face and finger), reducing MAE by 7.6% and 10.6% compared to using only face or finger data. Additionally, training on multiple indicators such as PPG and blood oxygen simultaneously reduces SpO2 estimation MAE by 17.8%. Repository: thuhci/SUMS Primary language: Python GitHub stars: 14 GitHub description: Summit Vitals: Multi-Camera and Multi-Signal Biosensing at High Altitudes

Overview: Kairos: The Context Keeper is an autonomous development supervisor powered by context engineering, designed to prevent AI context loss and ensure consistent development practices across projects. Key features: - **Context Engineering**: Advanced context preservation system that learns and adapts to your development patterns. - **Agent-Based Architecture**: Intelligent agents that handle different aspects of development autonomously. - **DevOps Integration**: Seamless integration with modern DevOps tools and cloud platforms. - **Memory Management**: Sophisticated knowledge graph that maintains project context across time. - **Adaptive Learning**: Self-improving system that gets better with each interaction. Tech stack: Python, FastAPI, Neo4j (Knowledge Graph), ChromaDB (Vector Store), Ollama, OpenRouter, Custom Model Routing. Why it matters: By leveraging advanced AI and memory management techniques, Kairos ensures consistent development practices and preserves critical project knowledge across time, helping teams deliver better software faster and prevent AI context loss in hackathons and developer onboarding.

Overview: Gemini Nano Banana 2.1 is an image model on APIMart, providing text-to-image and reference-image editing capabilities. It offers a single endpoint compatible with OpenAI, with a price of $0.0137 per image and supports 1K, 2K, and 4K outputs. Key features: - Single OpenAI-compatible endpoint for Gemini Nano Banana 2.1 - Supports text-to-image and reference-image editing - Per-image pricing of $0.0137 - Compatible with 1K, 2K, and 4K outputs - Asynchronous processing with task ID retrieval Tech stack: Python Why it matters: Gemini Nano Banana 2.1 simplifies access to a wide range of image models on APIMart, offering a single endpoint for text-to-image and reference-image editing at a cost-effective rate. This model is particularly useful for developers looking to integrate image processing capabilities into their projects without the need for multiple APIs. Repository: tz0cbrs/gemini-nano-banana-2-1-api-fr Primary language: Python GitHub stars: 9 GitHub description: Nano Banane / la Banane — Gemini Nano Banana 2.1 API (`gemini-nano-banana-2.1`): model id, pricing, parameters and copy-paste examples for the OpenAI-compatible APIMart endpoint.