Projects

Overview: This repository demonstrates how to convert Hugging Face tokenizers to ONNX models, enabling cross-language compatibility with embedding models in C#, Java, and Python. It provides dense embeddings without the need for third-party APIs or services, reducing latency and offering full control over the embedding pipeline. Key features: - Generate embeddings directly in C#, Java, or Python without third-party APIs or services - Reduced latency with local embedding generation - Full control over the embedding pipeline with no external dependencies - Works offline without internet connectivity requirements - Cross-platform compatibility Tech stack: ONNX Runtime Extensions Why it matters: This repository simplifies the process of using Hugging Face tokenizers in languages other than Python, addressing the challenge of implementing complex tokenizers from scratch or relying on Python interop. It provides a complete embedding pipeline in your preferred programming language, ensuring compatibility across multiple languages and environments.

Overview: This project is an implementation of AHD (Anchor-based History-stable Decoding) for Diffusion Large Language Models (dLLMs). It provides a training-free, plug-and-play dynamic decoding strategy for dLLMs. Key features: - AHD decoding implementation - Baseline (Semi-AR) decoding - Evaluation harness wrapper - Evaluation scripts for benchmarks like LLaDA, LLaDA-1.5, and MMADA Tech stack: Python Why it matters: This project offers a practical solution for dynamic decoding in dLLMs, improving performance without the need for extensive training. Repository: zs1314/AHD Primary language: Python GitHub stars: 39 GitHub description: [ACL 2026] Breaking Block Boundaries: Anchor-based History-stable Decoding for Diffusion Large Language Models

### Overview: This dataset contains traditional Chinese speech data collected from friends and YouTube videos, divided into three classes: positive, natural, and negative emotions. It includes sound speech data and spectrograms, designed for a virtual elderly companion agent. ### Key Features: - Sound speech data converted into spectrograms. - Three classes: positive, natural, and negative. - Train: 200 samples per class. - Test: 18 samples per class. ### Tech Stack: Python, TensorFlow, Keras. ### Why It Matters: This dataset is crucial for developing a virtual elderly companion agent, providing a comprehensive resource for training and testing models that can understand and respond to the emotions of the elderly.

### Overview: The **Segment Anything Model (SAM)** is a cutting-edge segmentation model that has been fine-tuned and applied to the challenging task of breast cancer tissue segmentation. This project utilizes Hugging Face's Transformers and fine-tunes the decoder to predict segmentation masks in medical imaging, specifically focusing on the Breast Cancer dataset. ### Key Features: - **Segment Anything Model (SAM)**: A versatile segmentation model capable of handling a wide range of tasks. - **Fine-Tuning**: The SAM model has been fine-tuned specifically for the task of breast cancer tissue segmentation. - **Medical Imaging**: The model is applied to medical images for accurate delineation of cancerous tissues. ### Tech Stack: - **Hugging Face Transformers**: Used for fine-tuning the SAM model. - **Segment Anything Model (SAM)**: A model that predicts segmentation masks based on prompts. ### Why It Matters: This project showcases the application of advanced machine learning techniques in medical imaging, specifically in the detection and segmentation of breast cancer tissues. The use of SAM for this task demonstrates the model's ability to handle complex and diverse segmentation tasks, making it a valuable tool in the field of medical imaging and cancer diagnosis.

### Overview: This project profile details the "Robotics and ROS - Learn by Doing! Manipulators" course, which is available on Udemy. The course is designed to guide students through the creation of a 3D printed robot arm controlled by voice using Amazon Alexa and ROS. The material covers concepts such as Gazebo Simulation, Robot Kinematics, ROS Basics, MoveIt!, and Arduino integration with ROS. ### Key Features: - 3D Printed robot arm - ROS-powered control - Alexa voice control - MoveIt! for robotic arm control - Arduino integration with ROS - Python and C++ code available ### Tech Stack: C++ ### Why it Matters: This course is ideal for students interested in robotics, particularly those looking to develop autonomous mobile robots. It covers essential topics in robotics, including ROS, Gazebo Simulation, and Arduino integration, providing a comprehensive learning experience. The ability to control a robotic arm with voice commands and integrate with Alexa makes the project both practical and engaging. ### GitHub Stars: 38 ### Repository: AntoBrandi/Robotics-and-ROS-Learn-by-Doing-Manipulators

Overview: This project aims to create a scalable and power-efficient Large Language Model (LLM) inference service using Kubernetes, with a custom power capping operator that accepts a Custom Resource Definition (CRD) to specify the power capping limit. The operator uses KEDA for autoscaling based on the specified power cap, and Kepler for monitoring power consumption. Key features: - Custom power capping operator for Kubernetes - Autoscaling based on power cap - Power monitoring using Kepler - Rack-level heating issues and rack-level cooling optimization Tech stack: Go Why it matters: This solution helps data centers efficiently utilize their compute resources while ensuring power and cooling constraints are not exceeded, reducing the risk of hardware damage, service disruptions, and operational costs. Repository: Climatik-Project/Climatik-Project GitHub stars: 38 GitHub description: Carbon Limiting Auto Tuning for Kubernetes

### Overview: Engram Memory is a high-scoring AI memory system that uses semantic embeddings and is self-hosted, free, and burns less tokens. It is designed to provide persistent memory across sessions without the need for API keys or cloud dependencies. ### Key features: - **Semantic Embeddings**: Engram uses semantic embeddings to store, search, recall, and forget memories. - **Self-Hosted**: The system runs on your own hardware, eliminating the need for API keys or cloud dependencies. - **Zero Tokens Burning**: Engram burns fewer tokens compared to other memory systems, making it more efficient. ### Tech stack: Qdrant, FastEmbed, and the MCP server. ### Why it matters: Engram provides a serious memory system that runs on your own hardware, costs nothing, and gets faster the more you use it, making it an ideal solution for AI agents that need persistent memory across sessions.

Overview: Vemcache is an in-memory vector database designed as a Redis equivalent for vector databases. Key features: - Supports vector database operations such as insertion, retrieval, removal, and similarity calculations. - Built on Rust, leveraging its performance and flexibility. - Built-in Docker support for easy deployment and isolation. - Detailed documentation available for advanced operations. Tech stack: Rust Why it matters: Vemcache simplifies the development of vector databases by providing a Redis-like interface for efficient vector operations, making it ideal for applications requiring high performance and scalability in vector-based data processing.

### Overview: stm32-speech-recognition-and-traduction is a project designed for the Operating Systems Design exam at the University of Milan (2020-2021). It features a speech recognition and translation system implemented on a STM32F407VG microcontroller, focusing on speech-to-text translation using pre-trained machine learning models. ### Key features: - **Speech Recognition**: Recognizes simple commands like turning LEDs on and off. - **Translation System**: Converts voice signals into text, displayed on the screen. - **Microcontroller**: Runs on STM32F407VG, suitable for limited memory and processing capacity. ### Tech stack: - **TensorFlow Lite** for speech recognition and translation. - **Google Colaboratory** for model training and inference. - **BSP** for microcontroller-specific drivers. - **PDM2PCM** for audio preprocessing. - **STM32_AI_AudioPreprocessing_Library** for audio processing. - **CMSIS DSP** for signal processing. - **X-CUBE AI** for AI-related functionalities. - **Picocom** for communication with the microcontroller. ### Why it matters: This project demonstrates the integration of machine learning with microcontroller-based systems, showcasing practical applications of operating systems in embedded systems.

### Overview: Light-MER is an open-source project designed to efficiently deploy multimodal emotion recognition models, specifically focusing on reducing the size of large emotion language models to sub-1B parameters. This project revisits generative multimodal emotion recognition (MER) from an efficiency perspective, aiming to replace a large 7B/8B model with a smaller, more powerful model. ### Key Features: - **Stage 1 SWD-H Distillation**: Transfers the multimodal emotion reasoning ability of a strong teacher into a sub-1B deployment model. - **Stage 2 M-GRPO Refinement**: Refines the model to improve performance and robustness. ### Tech Stack: - Python - PyTorch - Hugging Face Checkpoint ### Why It Matters: This project is significant because it demonstrates the feasibility of reducing the size of large emotion language models without sacrificing performance, potentially enabling more efficient and accessible emotion recognition systems.

Overview: This repository contains the codes for the paper "SORSA: Singular Values and Orthonormal Regularized Singular Vectors Adaptation of Large Language Models". It provides a novel PEFT method, SORSA, for efficient adaptation of large language models. Key features: - SORSA adapter consists of trainable principal singular weights and frozen residual weights. - These weights are initialized by performing SVD on pre-trained weights. - An orthonormal regularizer is implemented and analyzed. Tech stack: Python Why it matters: SORSA offers a significant improvement in parameter efficiency for fine-tuning large language models, making adaptation and fine-tuning more practical and efficient.

### Overview: StyleExpert is an innovative project that leverages the power of Style Experts to achieve diverse image stylization. This project is an official implementation of StyleExpert, a model presented at CVPR 2026. It uses a Mixture of Experts (MoE) architecture to achieve high-quality image stylization, making it particularly effective for complex semantic styles. ### Key Features: - **Style Representation Encoder:** Trained with InfoNCE loss to learn discriminative style features. - **MoE Fine-tuning:** Uses a similarity-aware gating mechanism to route styles to specialized LoRA experts. - **Dataset:** Provides the StyleExpert-40K dataset, containing 40,000 high-quality content-style-stylized triplets. ### Tech Stack: - Python - PyTorch with CUDA support ### Why It Matters: StyleExpert offers a powerful solution for diverse image stylization, making it ideal for applications requiring high-quality, semantically rich image transformations. Its effectiveness is demonstrated through its ability to achieve high-quality results on complex semantic styles, such as specific brushstrokes or materials. ### Repository: - **Repository:** HVision-NKU/StyleExpert - **Stars:** 38 - **Description:** Official implementation of StyleExpert (CVPR 2026) ### Dependencies and Installation: - **Python 3.10** - **PyTorch with CUDA support** - **Install requirements** via: ```bash pip install -r requirements.txt ``` ### Quick Inference: -

Overview: EEG-DIF is an algorithm designed to predict epileptic seizures using generative diffusion models and multi-channel EEG signals. Key features: - Clean version of the EEG-DIF algorithm for easy use with any EEG data. - Provides future predictions for any time frame using multi-channel EEG signals. - Uses Pytorch as the tech stack. Tech stack: Python. Why it matters: EEG-DIF offers a novel approach to early seizure warning through the use of generative diffusion models, potentially improving the accuracy and reliability of seizure prediction in the field of epilepsy research.

Overview: African Whisper is a framework designed to fine-tune and deploy the Whisper model on African languages, enhancing Automatic Speech Recognition (ASR) capabilities. Key features: - Fine-tuning of the Whisper model on various datasets - Metrics monitoring on Wandb - Production deployment for real-world applications - Model optimization using CTranslate2 - Word-level transcriptions and translations - Multi-speaker diarization - Improved alignment precision - Reduced hallucination Tech stack: Python Why it matters: African Whisper aims to advance ASR for African languages by providing robust and efficient tools for fine-tuning and deploying the Whisper model, ensuring high accuracy and clarity in transcription and translation. Repository: KevKibe/African-Whisper GitHub stars: 38 GitHub description: 🚀 Framework for seamless fine-tuning and deploying Whisper Model developed to advance Automatic Speech Recognition (ASR): translation and transcription capabilities for African languages.

### Overview: This Python script transcribes audio files to text using Google's speech recognition API. It supports multiple audio formats and languages, making it a versatile tool for transcription. ### Key Features: - Transcribes audio files from WAV, MP3, M4A, OGG, and FLAC formats. - Handles multiple languages including English, French, German, Spanish, Italian, Japanese, Korean, and Mandarin Chinese. - Supports the Google Speech Recognition API for accurate transcription. ### Tech Stack: Python ### Why It Matters: This script simplifies the process of converting audio to text, making it useful for researchers, journalists, and content creators who need to transcribe audio data efficiently.

### Overview: Advanced Optimizers is a comprehensive, all-in-one collection of state-of-the-art optimization algorithms designed for maximum efficiency, minimal memory footprint, and superior performance across diverse model architectures and training scenarios. ### Key Features: - **SinkSGD_adv:** A powerful new optimizer. - **Spectral Scaling:** Available across all optimizers, achieving width/rank invariant updates. - **Granular State Precision:** Modes for `state_precision` include `factored`, `fp32`, `bf16_sr`, and `int8_sr`. - **Factored Second Moment:** Available for all Adam variants. - **Variance Normalized Momentum:** Available for `AdamW_adv`, `SignSGD_adv`, and `SinkSGD_adv`. - **Universal Nesterov Momentum:** Replaces the hard-to-tune Simplified_AdEMAMix with Nesterov momentum. - **Preconditioning & Signs:** Added variance/confidence preconditioning (`snr_cond`) and adaptive stochastic sign (`stochastic_sign`). - **Improved CANS (`accelerated_ns`):** Enhanced for Muon variants. - **New OrthoGrad modes (`orthogonal_gradient`):** Standard OrthoGrad `flattened` and a new matrix-wise mode `iterative`. ### Tech Stack: Python ### Why It Matters: Advanced Optimizers offers a unified approach to optimization, ensuring efficient memory usage and superior performance across various deep learning models and training scenarios.

### Overview: MileBench is a benchmark designed to evaluate the capabilities of MLLMs in long context scenarios, comprising text and image data, and tasks requiring comprehension and generation. The repository contains evaluation code for the paper "MileBench: Benchmarking MLLMs in Long Context". ### Key Features: - **Diagnostic Evaluation**: Focuses on long-context recall abilities using tasks like needle-in-a-haystack and image retrieval. - **Realistic Evaluation**: Tests the model in a realistic manner, including temporal multi-image tasks and semantic multi-image tasks. - **Model Configuration**: Supports customization of model configurations, including model name, directory, and generation parameters. - **Environment Setup**: Requires installation of specific packages and setting up the environment for evaluation. ### Tech Stack: Python, PyTorch, transformers, accelerate ### Why it Matters: The results highlight the current limitations of open-source MLLMs in handling long-context tasks, emphasizing the need for increased research focus on such capabilities.

### Overview: TimeDiff is a code repository designed for generating realistic synthetic healthcare data using diffusion models. This project aims to provide privacy-preserving synthetic electronic health record (EHR) time series data, which is crucial for research and development in the healthcare domain. ### Key Features: - **Code Implementation**: The repository contains the code for generating synthetic EHR time series data using a diffusion model. - **Model Training**: The script `etdiff_train.py` initializes model training, generating synthetic data automatically after completion. - **Data Pre-processing**: The `preprocess` directory contains Jupyter notebooks and Python scripts for preprocessing EHR data. - **Datasets**: The repository includes datasets from the MIMIC-III, MIMIC-IV, HIRID, and Physionet repositories. ### Tech Stack: The project uses Jupyter Notebook as the primary language. ### Why It Matters: TimeDiff is significant for researchers and developers in the healthcare industry who need to generate realistic synthetic data for training and testing models without compromising patient privacy. It provides a reliable and privacy-preserving solution for various applications in healthcare research and development. ### GitHub Stars: The repository has received 37 GitHub stars, indicating its popularity and usefulness in the community. ### Repository: The repository is named `MuhangTian/TimeDiff` and is hosted on GitHub. ### Primary Language: The primary language used in the repository is Jupyter Notebook. ### GitHub Description: The repository contains the code implementation for TimeDiff, a diffusion model designed for generating mixed-type time

Overview: AsterVec is an embedded on-device vector database optimized for low memory footprint, designed to be used with AI agents and desktop RAG systems. It stores its index on disk and minimizes memory usage by keeping only the necessary data in RAM. Key features: - Minimizes memory usage by keeping index on disk and only minimal data in RAM. - Assumed data growth, so the index can absorb new vectors as they arrive. - Thread-safe Python and C++ library for easy integration into applications. - Supports inserts, updates, and deletes in place. - Fast initial load with bulk_build for large datasets. - Measures 2.2x less memory, 5.2x more queries per second, and 7.3x more inserts per second compared to other engines. Tech stack: C++/Python, optimized for low memory footprint. Why it matters: AsterVec is designed to help developers minimize memory usage, handle growing data efficiently, and provide fast, accurate search capabilities without requiring a separate service or large amounts of RAM.

Overview: This ROS package aims to bring offline speech recognition capabilities to the ROS ecosystem, building upon the popular pocketsphinx library. Key features: - ROS bindings for pocketsphinx - Offline speech recognition capabilities Tech stack: Python Why it matters: This package extends the functionality of pocketsphinx, making it more accessible and useful in the ROS environment, thereby enhancing the capabilities of speech recognition in robotics applications.