Projects

Overview: This project aims to enable robots to perform long-horizon dexterous tasks with imitation learning. Key features: - Data-Efficient Imitation Learning - Robotic Assembly - Custom Fork of FurnitureBench for Data Collection Tech stack: Jupyter Notebook Why it matters: This project provides a robust framework for robots to learn from human demonstrations and perform complex tasks efficiently, making it invaluable in various industrial applications. Repository: ankile/imitation-juicer GitHub stars: 44

Overview: SeaTunnel Tools is a collection of auxiliary tools designed to enhance developer and operator productivity in Apache SeaTunnel, a high-performance, distributed, massive data integration tool. Key features: - SeaTunnel Skill: A Claude AI integration for SeaTunnel operations, providing instant help with documentation and best practices. - SeaTunnel MCP Server: A Model Context Protocol for LLM integration, facilitating seamless integration with large language models. - x2seatunnel: A configuration converter for DataX to SeaTunnel format, streamlining configuration processes. Tech stack: Java Why it matters: SeaTunnel Tools simplifies the integration and management of data across various sources and destinations, making it easier for developers to handle large-scale data processing and analysis. Repository: apache/seatunnel-tools Primary language: Java GitHub stars: 44 GitHub description: Apache SeaTunnel is a multimodal, high-performance, distributed, massive data integration tool.

### Overview: HairPort is a 3D-aware hairstyle transfer framework designed for in-context bald conversion, pose-aligned hair import, and identity-preserving synthesis. It leverages 3D-aware alignment to handle large pose and scale differences, ensuring accurate and contextually relevant hairstyle transfer. ### Key features: - **In-Context Bald Conversion:** Seamlessly converts a bald hairstyle to a source face. - **Pose-Aligned Hair Import:** Imports hair from a source face while maintaining the original pose. - **Identity-Preserving Synthesis:** Synthesizes the final hairstyle while preserving the identity of the source face. ### Tech stack: Python ### Why it matters: HairPort addresses the limitations of existing hairstyle transfer methods by explicitly handling large pose and scale differences, making it a significant advancement in the field of computer graphics and visual effects. It enables users to explore new hairstyles without physically altering their hair, with applications in virtual try-on systems, augmented reality, and entertainment.

### Overview: This collection features 252 curated FLUX 3 Image Klein prompts across 9 categories, providing preview images for easy selection. ### Key Features: - **Portrait & Avatar**: Prompts for professional-looking portraits and headshots. - **Product & E-commerce**: Images for product showcases and e-commerce marketing. - **Poster & Typography**: Designs for promotional posters and typography. - **Infographic & Diagram**: Visual representations for data and information. - **Anime & Illustration**: Artwork for anime and illustration projects. - **3D & Material**: Models and textures for 3D design. - **Food & Lifestyle**: Images for food and lifestyle photography. - **Architecture & Interior**: Designs for architectural and interior projects. - **Nature & Landscape**: Scenery and landscape images. ### Tech Stack: Utilizes open-weight FLUX 3 with layout control for creative prompts. ### Why It Matters: Provides a curated set of prompts that work well, allowing developers to quickly find and use high-quality images for their projects without having to create them themselves.

Overview: "Whisper-LM" is an extension to OpenAI's Whisper models that integrates n-gram and large language models (LLM) to enhance automatic speech recognition (ASR) performance, particularly for low-resource languages. Key features: - Integrates n-gram and LLM support - Improves ASR performance for low-resource languages - Compatible with Hugging Face's Whisper model format Tech stack: Jupyter Notebook Why it matters: This repository provides a practical solution for improving ASR models in low-resource languages, making it easier to fine-tune Whisper models and achieve better transcription results.

### Overview: CustomNeRF is a PyTorch implementation of a novel 3D scene editing model designed for adaptive source-driven 3D scene editing. It unifies text descriptions or reference images as editing prompts, addressing the challenges of accurately editing foreground regions and maintaining multi-view consistency in single-view reference images. ### Key Features: 1. **Local-Global Iterative Training (LGIE):** Alternates between foreground region editing and full-image editing to ensure precise foreground manipulation while preserving the background. 2. **Class-Guided Regularization:** Exploits class priors within the generation model to alleviate inconsistency among different views in image-driven editing. 3. **Versatile Dataset Support:** Compatible with popular NeRF datasets and custom datasets extracted from your own camera. ### Tech Stack: - Python ### Why It Matters: CustomNeRF offers a powerful tool for 3D scene editing by providing precise and adaptable solutions for both text-driven and image-driven editing tasks, making it a valuable addition to the field of 3D scene editing and rendering.

Overview: This project is a privacy permission tool for iOS applications, including the ability to check and manage permissions for various services and features such as location services, contacts, calendars, reminders, photos, Bluetooth, microphone, camera, health, media, and notifications. It also includes the registration and judgment of notifications. Key features: - Checks and manages privacy permissions for location services, Bluetooth, microphone, camera, and notifications. - Registers and manages location permissions. - Provides options for handling permission prompts, such as prompting users to grant permissions or only showing permission requests. - Supports both push and local notification registration. - Configures privacy settings for health and HomeKit. - Requires certificates for HomeKit and media access. Tech stack: Objective-C Why it matters: This tool simplifies the process of managing privacy permissions in iOS applications, ensuring that users are informed and can easily grant or deny permissions as needed. It provides a comprehensive solution for developers looking to enhance user experience by respecting privacy settings.

Overview: This project allows users to find compatible GPUs for running AI models, quantization, VRAM, and speed ranges, along with executable commands. It provides detailed information on 3 AI models, 350 AI models, and 160 GPU presets. Key features: - Compatible AI models - GPU presets - VRAM and speed ranges - Executable commands Tech stack: JavaScript Why it matters: This project helps developers find the best GPU for their AI projects, ensuring compatibility with various models and performance specifications. Repository: jaeseok614/ai-hardware-fit GitHub stars: 44 GitHub description: Can my GPU run this? Pick a GPU to get 3 fitting AI models, quantization, VRAM, speed ranges, and run commands.

Overview: SLEB is an official implementation of the ICML 2024 paper that introduces a novel approach to streamline Large Language Models (LLMs) by eliminating redundant transformer blocks. This method aims to enhance the processing speed of LLMs while maintaining superior perplexity and accuracy. Key features: - Streamlines LLMs by removing redundant transformer blocks. - Meets the goal of accelerating LLM inference while maintaining superior perplexity and accuracy. - Integrates with the EuletherAI's lm-evaluation-harness for zero-shot task performance evaluation. Tech stack: Python Why it matters: SLEB offers a promising technique for enhancing the efficiency of LLMs by reducing their size and complexity, making them more practical for real-world applications.

### Overview: FigGen is a text-to-figure generation model that converts textual descriptions from research papers into scientific figures. It leverages the power of diffusion models and Bert transformers to create high-quality visual representations of scientific papers. ### Key features: - **Text-to-Image Generation:** Converts textual descriptions into scientific figures. - **Latent Diffusion Model:** Uses a latent diffusion model to generate figures conditioned on text. - **Text-to-Image Transformer:** Jointly trains a transformer to learn text embeddings and text-to-image generation. ### Tech stack: - Python ### Why it matters: FigGen is a significant advancement in the field of scientific visualization, enabling researchers to quickly generate figures directly from their research papers, enhancing the reproducibility and accessibility of scientific findings.

### Overview: VisBridge MCP is a token-efficient, provider-agnostic vision layer designed for MCP clients. It provides a single `analyze_image` tool that allows any MCP client to see, describe, OCR, or inspect images through an OpenAI-compatible vision provider. ### Key Features: - **Two Transports:** Supports local clients using stdio (Claude Code, Cursor) and Streamable HTTP for clients that connect over HTTP. - **Three Modes:** Supports three modes: `describe` (summarize), `ocr` (extract text), and `inspect` (answer a specific question via `prompt`). - **Detail: Low by Default:** Offers cheaper OCR than `high` mode. - **Fence-Free OCR:** Strips markdown code fences defensively. - **Four Image Sources:** Supports local `path`, `url`, `base64` (data URL or raw), and `resource` (resolved against a configured directory). - **Hardened SSRF Protection:** Blocks private/loopback/link-local IPs, resolves DNS before fetch, and limits MIME types and sizes. - **Local Preprocessing:** Downscales oversized images and uses the first frame for animated GIFs, with a `warnings` entry noting every resize. - **Structured Errors:** Uses a 10-code taxonomy and MCP-safe messages. - **Observability:** Provides token/cost estimation and logs one JSON line per call to stderr. ### Tech Stack: Node.js >= 20 ### Why It Matters: VisBridge MCP aims to give a model without native vision or any client that wants

Overview: This project provides a set of packages for integrating speech and voice inputs into Emacs, enhancing the user experience with natural language interactions. Key features: - Speech input integration - Voice input recognition - Voice command support Tech stack: C Why it matters: This project offers a valuable tool for Emacs users looking to enhance their productivity and accessibility by integrating speech and voice inputs, making interactions with the editor more intuitive and efficient.

### Overview: 该仓库包含多模态情感分析实验的配套代码，用于处理文本和图像数据，构建多模态模型，并进行实验和预测。 ### Key features: - 多模态情感分析模型构建 - 训练数据包含文本和图像 - 输出预测结果 - 可调整参数查看功能 ### Tech stack: Jupyter Notebook ### Why it matters: 该实验提供了多模态情感分析的完整解决方案，适用于文本和图像数据的处理和分析，具有实际应用价值。

### Overview: This repository provides a comprehensive framework for preprocessing and training LSTM networks for phoneme classification on the TIMIT dataset. The project aims to achieve high accuracy by generating and normalizing audio files and training a model that predicts phonemes in the middle of each interval. ### Key Features: - **Data Preprocessing**: The dataset is modified to include phoneme classes, resampling audio, and creating MLF files. - **Model Training**: LSTM networks are trained to classify phonemes with a focus on generating predictions in phoneme intervals. - **Performance**: The model achieves 82% accuracy on the TIMIT dataset, comparable to results from [Graves et al (2013)](https://arxiv.org/abs/1303.5778). ### Tech Stack: - **Python**: The primary programming language used for the project. - **RNN**: Recurrent Neural Networks are employed for the phoneme classification task. ### Why It Matters: This project offers a robust solution for speech recognition, particularly useful for researchers and developers interested in improving speech recognition models on the TIMIT dataset.

Overview: MedTok is a multimodal medical code tokenizer designed for processing and tokenizing medical codes from Electronic Health Records (EHRs). It integrates text descriptions and relational context to improve tokenization accuracy and clinical reasoning. Key features: - Multimodal tokenization of medical codes - Uses language model encoder for text and graph encoder for relational context - Preserves modality-specific and cross-modality information - Integrates into five EHR models for improved performance Tech stack: Python Why it matters: MedTok significantly improves tokenization accuracy in EHR models, leading to better outcomes in clinical tasks such as prediction, classification, and recommendation. It also demonstrates potential for integration with medical QA systems. Repository: mims-harvard/MedTok Primary language: Python GitHub stars: 44 GitHub description: [ICML'25] MedTok: Multimodal Medical Code Tokenizer

### Overview: This project aims to develop a speech recognition module for Red Hen Lab's audio processing pipeline. The initial goal is to extend the current Deep Speech model to handle audio-only speech recognition. The next phase involves developing an Audio-Visual Speech Recognition (AVSR) system by incorporating visual modalities from news videos. ### Key Features: - **Audio-only Speech Recognition**: Extend the Deep Speech model for speech recognition using audio-only data. - **Audio-Visual Speech Recognition (AVSR)**: Develop a multi-modal Speech to Text system by extracting visual modalities from news videos and concatenating them with the audio inputs. ### Tech Stack: - Python ### Why It Matters: This project is significant for Red Hen Lab as it aims to enhance their speech recognition capabilities, particularly for integrating visual information into the recognition process. The development of AVSR can lead to more accurate and comprehensive speech recognition systems, which is crucial for applications such as lip reading and multimedia content understanding. ### Repository: pandeydivesh15/AVSR-Deep-Speech - **GitHub Stars**: 44 - **GitHub Description**: Google Summer of Code 2017 Project: Development of Speech Recognition Module for Red Hen Lab ### Primary Language: Python - **GitHub Description**: Google Summer of Code 2017 Project: Development of Speech Recognition Module for Red Hen Lab ### GitHub Description: This project is based on the approach discussed in the paper "Deep Speech" (arXiv:1412.5567). The aim is to extend the Deep

Overview: qwen3-rs is an educational Rust project designed to explore and run Qwen3 language family models, focusing on clear, modular, and approachable code with minimal dependencies. Key features: - Tokenization - Quantization - Sampling - Efficient inference using memory mapping and zero-copy techniques Tech stack: Rust Why it matters: This project provides a comprehensive educational resource for understanding transformer architectures, quantization, and efficient inference in Rust, making it ideal for learners and developers interested in the Qwen3 language family.

Overview: This repository contains code used for exploring various territories of pipeline-model-parallelism, including partitioning, training, and fine-tuning of large neural networks. It supports models like Huggingface Transformers and Torchvision, and includes a study on pipeline staleness mitigation. Key features: - Partitioning models - Training and fine-tuning of large neural networks - Support for models like Huggingface Transformers and Torchvision - Pipeline staleness mitigation study Tech stack: Python Why it matters: This repository is a valuable resource for researchers and developers interested in fine-tuning neural networks on commodity hardware, particularly focusing on pipeline model parallelism. Repository: saareliad/FTPipe GitHub stars: 44 GitHub description: FTPipe and related pipeline model parallelism research.

Overview: This Python-based solution efficiently solves Google's reCAPTCHA v2 challenges, offering both synchronous and asynchronous operations with API endpoints for easy integration. It uses an audio challenge method with speech recognition for reliable solving, supporting multiple operation modes and proxy support. Key features: - Supports synchronous, asynchronous, and API-based solving - Uses speech recognition for automatic solving of audio challenges - Built-in proxy support to avoid rate limits - Batch processing for multiple CAPTCHAs - Comprehensive debug logs for troubleshooting - Automatic rate limit detection and handling - Enterprise support for both standard and enterprise reCAPTCHA - Multiple methods for reliable token extraction - Robust error handling with detailed feedback Tech stack: Python Why it matters: This script provides a fast and reliable solution for solving reCAPTCHA v2 challenges, making it ideal for developers looking to integrate security measures in their projects efficiently and securely. Repository: sexfrance/RecaptchaV2-Solver Primary language: Python GitHub stars: 44 GitHub description: A Python-based solution for solving Google's reCAPTCHA v2 challenges efficiently (3-12 seconds solve time). The script supports both synchronous and asynchronous operations, with API endpoints for easy integration. Uses audio challenge method with speech recognition for reliable solving.

### Overview: Portable Docker APP is a collection of Docker-based applications designed to simplify and enhance various tasks, including vector database, image search, and text analysis. It aims to provide a lightweight and portable solution for developers and enthusiasts. ### Key Features: - **2024**: Out-of-the-box vector database, ML/AI first. - **2023**: Reliable Image Search Engine. - **2022**: Text Emotion / Sentiment Analysis Tools. ### Tech Stack: Python-based applications using Docker. ### Why It Matters: Portable Docker APP simplifies the deployment and management of complex applications, making them accessible and easy to use for developers and enthusiasts.