Projects

### Overview: FASHN VTON v1.5 is an efficient maskless virtual try-on model that generates photorealistic images directly in pixel space without requiring segmentation masks. This technology allows users to see how a garment would look on a person's face or body, enhancing the virtual try-on experience. ### Key Features: - **Efficient Pixel Space Generation**: The model generates images in pixel space, eliminating the need for segmentation masks. - **Supports Multiple Categories**: The model supports upper body (tops), lower body (bottoms), and full body (one-pieces). - **Versatile Inputs**: The model can take both model photos and flat-lay product shots as garment inputs. ### Tech Stack: The project uses Python and ONNX Runtime for inference, with additional support for GPU acceleration for better performance on RTX 30xx/40xx, A100, and H100 GPUs. ### Why It Matters: This technology simplifies the virtual try-on process by providing a more intuitive and efficient way to visualize how clothing would fit and look on a person, enhancing the overall user experience in fashion applications.

Overview: This project is a voice activity detection system utilizing deep learning techniques with TensorFlow. It analyzes audio data to identify when a person is speaking, which is crucial for applications such as call center monitoring, security systems, and more. Key features: - Implements advanced deep learning models for accurate detection - Utilizes TensorFlow for efficient model training and inference - Integrates with various audio input devices Tech stack: Python Why it matters: This project is essential for applications requiring real-time monitoring of audio inputs, ensuring timely alerts and improved security measures.

Overview: Create-LLM is a fast and efficient CLI tool designed to streamline the process of building and training language models. It scaffolds production-ready PyTorch projects in seconds, making it an ideal choice for developers looking to quickly prototype or deploy language models. Key features: - Right-sized templates for different use cases (NANO, TINY, SMALL, BASE) - Complete toolkit with PyTorch infrastructure, data preprocessing, tokenizer training, checkpoint management, and more - Smart defaults that auto-detect vocab size, handle sequence length mismatches, and suggest optimal hyperparameters - Plugin system for optional integrations like WandB and HuggingFace - Quick start options including one-command setup and interactive setup Tech stack: TypeScript Why it matters: Create-LLM simplifies the complex process of training language models, allowing developers to focus on the model architecture and data without getting bogged down in the technical details. Its fast setup and comprehensive toolkit make it an invaluable tool for anyone looking to quickly prototype or deploy language models.

### Overview: This repository contains a comprehensive list of papers that explore the application of large language models (LLMs) in recommendation systems. The papers address various aspects of recommendation, including the use of foundation models, the necessity of ID embedding, the shift from matching to generating paradigms, and the potential of LLMs in enhancing recommendation systems. The papers also discuss datasets and benchmarks for transferable or multimodal recommendation systems. ### Key features: - **Foundation Models for Recommender Systems**: Papers exploring the use of foundation models in recommendation systems, focusing on their effectiveness and potential integration with large language models. - **ID Embeddings and Relevance**: Discussions on the importance of ID embeddings in recommendation systems and their potential replacement with LLMs. - **Transfer Learning and Multimodal Recommendation**: Papers examining the transfer learning approach for recommendation systems and the integration of multimodal data for better performance. - **Large Language Models for Recommendation**: Papers detailing the use of large language models in recommendation systems, including their impact on matching and generating paradigms. - **Datasets and Benchmarks**: A collection of datasets and benchmarks for evaluating and benchmarking recommendation systems, including image, text, and multi-domain datasets. ### Tech stack: Python, Jupyter Notebooks, and various libraries for data processing and visualization. ### Why it matters: This repository is crucial for researchers and developers looking to understand the current state of research in recommendation systems and the potential of large language models in enhancing these systems. It provides a valuable resource for those interested in the intersection of recommendation, large language models, and the future

Overview: CBraMod is a novel EEG decoding model designed for various clinical and brain-computer interface (BCI) applications. It leverages a criss-cross brain foundation to enhance decoding accuracy. Key features: - A novel criss-cross brain foundation model for EEG decoding. - Pretrained on a pretraining dataset or a custom dataset. - Available on Hugging Face for fine-tuning. - Supports both pretraining and finetuning on downstream datasets. Tech stack: Python, PyTorch, Hugging Face. Why it matters: CBraMod offers improved EEG decoding capabilities, making it a valuable tool for researchers and practitioners in the field of brain-computer interaction and clinical applications.

### Overview: This repository contains the official implementation of the paper "Revisiting Deep Learning Models for Tabular Data" presented at NeurIPS 2021. The paper focuses on architectures for tabular data problems, and the results show that a simple tuned MLP is still a good baseline, and the ResNet model further highlights this point. The Transformer architecture, FT-Transformer, is introduced as a new adaptation of the Transformer architecture for tabular data problems, which significantly outperforms other deep models on benchmarks and reduces the gap between GBDT and DL models on datasets dominated by GBDT. ### Key features: - The paper focuses on architectures for tabular data problems. - The results show that a simple tuned MLP is still a good baseline, and the ResNet model further highlights this point. - The Transformer architecture, FT-Transformer, is introduced as a new adaptation of the Transformer architecture for tabular data problems, which significantly outperforms other deep models on benchmarks and reduces the gap between GBDT and DL models on datasets dominated by GBDT. ### Tech stack: - Python ### Why it matters: The paper demonstrates that simple and effective baseline models can still perform well on tabular data problems, and the Transformer architecture, FT-Transformer, offers a new way to adapt the Transformer architecture for tabular data problems, significantly improving performance compared to other deep models on benchmarks and reducing the gap between GBDT and DL models on datasets dominated by GBDT. ### Repository: The repository is located at `https://github.com/yandex-research

### Overview: Program that enables users to ask questions about their documents and audio or video files, providing more reliable responses from an LLM. ### Key features: - Vector database integration for efficient querying of documents and multimedia files. - Support for various file types including PDF, DOCX, TXT, HTML, HMTL, MD, CSV, XLS, XLSM, RTF, EML, MSG, PNG, JPG, BMP, GIF, TIF, TIFF, MP3, WAV, M4A, OGG, WMA, FLAC. - Retrieval augmented generation, commonly referred to as "retrieval augmented generation." ### Tech stack: Python ### Why it matters: This program allows users to seamlessly integrate vector databases into their document and multimedia management workflows, enhancing the reliability of AI-generated responses.

Overview: This project is an AI-driven YouTube automation pipeline that generates educational videos and uploads them to the channel using a combination of large language models and Python scripts. It automates the entire process from script writing to video creation and upload, ensuring zero human intervention. Key features: - Gemini 2.5 Flash for AI script generation - YouTube Short videos for daily uploads - Automatic script updates to the content plan - Pexels stock imagery for dynamic visuals - Professional audio narration with background music - Custom thumbnails for each video - Self-updating repository for continuous content management Tech stack: Python Why it matters: This project showcases the potential of AI in automating repetitive tasks and creating engaging educational content, making it a valuable tool for anyone looking to automate their YouTube channel without the need for manual intervention.

Overview: OpenGPT is a framework designed to create grounded instruction-based datasets and train conversational domain expert Large Language Models (LLMs). It is particularly useful for healthcare applications, offering medical datasets and a conversational model trained using OpenGPT. Key features: - Grounded instruction-based datasets for healthcare - Conversational model trained using OpenGPT - Available datasets: NHS UK Q/A, NHS UK Conversations, and Medical Task/Solution - Installation guide using pip - Tutorials for creating mini conversational LLMs and training LLMs Tech stack: Jupyter Notebook Why it matters: OpenGPT simplifies the process of creating and training LLMs for healthcare applications, providing ready-to-use datasets and a conversational model that can be easily integrated into healthcare systems. Repository: CogStack/OpenGPT GitHub stars: 371 GitHub description: A framework for creating grounded instruction based datasets and training conversational domain expert Large Language Models (LLMs).

### Overview: UniGoal is a unified graph representation for zero-shot goal-oriented navigation, enabling direct application across various scenes and goals without training. This innovative approach was accepted to CVPR 2025 and has garnered significant interest and support from the community. ### Key Features: - **Unified Graph Representation**: A single graph-based framework that can handle zero-shot navigation. - **Zero-Shot Capability**: Directly applicable to different scenes and goals without any training. - **Scenes and Goals**: Supports various scenes and goals, making it versatile and adaptable. - **Real-World Deployment**: Demonstrated in real-world scenarios and simulation environments, showcasing practical applications. - **Simulation Environment**: Available for easy testing and validation of the method. ### Tech Stack: - **Python**: The primary programming language used for implementation and development. - **Habitat-Sim**: A simulator framework for training and testing agents in complex environments. - **Habitat-Lab**: A library for building and training agents in the Habitat simulator. - **Segment Anything Model**: A state-of-the-art model for instance segmentation. - **GroundingDINO**: A model for object detection and segmentation in images. ### Why It Matters: - **Versatility**: The method's ability to handle zero-shot navigation across different scenes and goals makes it highly versatile and adaptable. - **Ease of Use**: The unified graph representation simplifies the integration of different navigation scenarios, making it easier to implement and deploy. - **Practical Applications**: The demonstrated real-world and simulation environments highlight the method's potential for practical use in

Overview: LiLT is an official PyTorch implementation of the LiLT paper, which is designed for structured document understanding. It leverages pre-trained models to handle language-independent layout tasks. Key features: - Pre-trained on a single language (English) and fine-tuned on other languages using corresponding monolingual/multilingual pre-trained models. - Supports fine-tuning on various document understanding tasks like semantic entity recognition. - Available pre-trained checkpoints for English (lilt-roberta-en-base) and multiple languages. Tech stack: Python Why it matters: LiLT simplifies the process of applying language-independent layout transformers for structured document understanding, making it accessible for researchers and practitioners without extensive language modeling expertise. Repository: jpWang/LiLT GitHub stars: 371 GitHub description: Official PyTorch implementation of LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding (ACL 2022)

Overview: Segformer is an efficient neural network for semantic segmentation, implemented in PyTorch. It utilizes attention and MLP for improved performance. Key features: - Efficient attention mechanism - Multi-stage architecture - Supports multiple input sizes - Compatible with PyTorch framework Tech stack: PyTorch Why it matters: Segformer offers a significant improvement in segmentation accuracy and efficiency, making it a valuable tool for researchers and practitioners in computer vision and image processing. Repository: lucidrains/segformer-pytorch Primary language: Python GitHub stars: 371 GitHub description: Implementation of Segformer, Attention + MLP neural network for segmentation, in Pytorch.

Overview: This repository contains the code for the Gazelle Joint Speech Language Model, a model designed to respond directly to audio. Key features: - Joint speech-language model - Inference code heavily derived from Huggingface's Llava implementation - Available checkpoints: v0.2, v0.2-dpo, v0.1 Tech stack: Python Why it matters: This model is useful for applications requiring real-time audio response, such as voice assistants or interactive voice recognition systems. Repository: tincans-ai/gazelle Primary language: Python GitHub stars: 371 GitHub description: Joint speech-language model - respond directly to audio!

Overview: This repository contains a series of hands-on notebooks designed to teach the fundamentals of Natural Language Processing (NLP) from first principles. It aims to provide a solid understanding of NLP techniques and concepts, emphasizing theoretical aspects and practical implementation using various tools and libraries. Key features: - Tokenization - Lemmatization - Stemming - Sentence Segmentation Tech stack: Jupyter Notebook Why it matters: Understanding the fundamentals of NLP is crucial for developing innovative techniques, improving research, and creating safe, human-centered AI and NLP technologies that affect society at large. GitHub stars: 370 Repository: dair-ai/nlp_fundamentals

Overview: This repository contains large-coverage, machine-readable lemma/token pairs in 23 languages, collected from various sources and used for query expansion in fulltext searches. Key features: - 23 languages - 370 GitHub stars - Available under the Open Database License Tech stack: Plain text files (zipped) with UTF-8 encoding and Windows-style line breaks. Why it matters: These lemma/token pairs are useful for expanding search queries, improving search relevance, and enhancing the accuracy of fulltext searches in various languages.

Overview: This repository contains the PyTorch implementation of a graph convolutional neural network (GCN) designed for relation extraction using pruned dependency trees. The model improves performance by leveraging dependency structures to bridge remote words, outperforming sequence models and recursive models like TreeLSTM. Key features: - Graph Convolution over Pruned Dependency Trees - Path-centric Pruning to eliminate irrelevant information - Improved performance for long-range relations compared to sequence models - Efficient and parallelizable compared to recursive models Tech stack: Python Why it matters: This GCN model enhances relation extraction by effectively managing dependency structures, leading to better performance and faster training compared to traditional sequence models and recursive models.

Overview: This project is a Spring AI-powered expert system demo that showcases how to integrate AI capabilities into a Java application using Spring AI, including access to terms and conditions via RAG, the ability to call Java methods, and interaction with a large language model (LLM) through a user interface. Key features: - Access to terms and conditions via RAG - Ability to call Java methods - Interaction with a large language model (LLM) through a user interface Tech stack: Spring AI Why it matters: This project demonstrates the integration of AI into a Java application, providing a practical example of how to build an AI-powered system that interacts with user inputs and can perform actions through Java methods, while also leveraging a large language model to provide insights and responses. Repository: tzolov/playground-flight-booking Primary language: Java GitHub stars: 370 GitHub description: Spring AI powered expert system demo

Overview: This repository contains scripts for fine-tuning and evaluating Whisper models for Automatic Speech Recognition (ASR) using custom datasets or datasets from the Hugging Face library. Key features: - Fine-tuning and evaluation of Whisper models on custom datasets - Fine-tuning and evaluation of Whisper models on datasets from the Hugging Face library - Support for custom datasets preparation - Support for hyperparameter tuning - Support for fine-tuning on a dataset from the Hugging Face library - Support for fine-tuning on a custom dataset - Support for evaluating on a dataset from the Hugging Face library - Support for evaluating on a custom dataset - Support for transcribing a single audio file - Support for faster evaluation using Whisper Jax - Support for extracting embeddings from Whisper models with various configurations Tech stack: Python Why it matters: This repository provides a comprehensive solution for fine-tuning and evaluating Whisper models for ASR, making it a valuable resource for developers working on speech recognition projects.

Overview: Material Anything is a cutting-edge model that generates PBR materials for any 3D object, including texture-less, albedo-only, generated, and scanned objects. This project is a significant contribution to the field of 3D rendering and material synthesis. Key features: - Texture-less materials - Albedo-only materials - Generated materials - Scanned materials Tech stack: Python Why it matters: Material Anything simplifies the process of generating high-quality PBR materials for various 3D objects, making it easier for developers and artists to create realistic textures and materials for their projects. This tool is particularly useful in fields such as game development, virtual reality, and architectural visualization, where precise material representation is crucial. Repository: 3DTopia/MaterialAnything GitHub stars: 369

Overview: TextDescriptives is a Python library designed to calculate a wide range of text metrics using spaCy, providing a comprehensive suite of tools for text analysis. Key features: - Extracts and calculates various metrics from text data. - Supports multiple languages and spaCy models. - Provides a web app for extracting metrics without code. - Updated API with new components and improved documentation. Tech stack: Python, spaCy v.3 pipeline components, spaCy models. Why it matters: Ideal for developers looking to automate text analysis and extract meaningful metrics quickly and easily, making it a valuable tool for various applications in natural language processing and text mining.