Projects

### Overview: AI WorkDeck is an open-source AI workspace designed to streamline professional document work, integrating legal practice into the workflow. ### Key features: - **Document Editor with AI Integration**: Works alongside the document editor, project files, source references, and version history. - **AI-Generated Changes**: Agents assist in reviewing and accepting or rejecting AI-generated changes. - **Project History and Collaboration**: Keeps track of project history and allows developers to build on the same editor, agent tools, and plugin runtime. ### Tech stack: JavaScript ### Why it matters: AI WorkDeck provides a shared foundation for professional AI, integrating demanding legal workflows and enabling specialists to build on top of the same editor, agent tools, and plugin runtime.

### Overview: eShopLite is a set of reference .NET applications showcasing various advanced features such as Semantic Search, Model Context Protocol (MCP), Reasoning models, and vector databases. It is built using .NET and Aspire, providing a stable .NET 10 baseline and an experimental .NET 11 RC2 scenario for Blazor AI components. ### Key Features: - **Semantic Search**: Implements an eCommerce site with keyword and semantic search capabilities. - **Model Context Protocol (MCP)**: Server and client implementations for integrating AI models. - **Vector Database**: In-memory, Azure AI Search, and Chroma DB implementations. - **Real-time Audio**: Enhanced search features powered by the GPT-4o Realtime Audio API. ### Tech Stack: .NET This project is built using the .NET framework, leveraging Aspire for its modern application architecture. ### Why It Matters: - **Scenarios**: Demonstrates various advanced search functionalities and AI integration. - **Stability**: Baseline compatible with .NET 10 and experimental .NET 11 RC2. - **Scalability**: Supports real-time audio and vector database implementations. - **Integration**: Integrates with multiple AI models and vector databases. This project is ideal for developers looking to implement advanced search and AI functionalities in their .NET applications.

Overview: SyncDiffusion is an official implementation of a method for generating coherent montages using synchronized joint diffusions, presented at NeurIPS 2023. Key features: - Synchronized joint diffusions through gradient descent from perceptual similarity loss - Plug-and-play module for seamless montage generation - Conditional generation based on ControlNet (Zhang et al.) - Released to NeurIPS 2023 Tech stack: Jupyter Notebook Why it matters: SyncDiffusion offers a significant improvement in coherence and fidelity compared to previous methods, achieving 66.35% in user studies while maintaining high quality and compatibility with input prompts.

### Overview: YOLOFuse 是一个基于 Ultralytics YOLO 框架的增强型目标检测系统，专为多模态感知任务设计。它创新性地引入双流处理架构，支持 RGB 与红外（IR）图像的协同分析与特征融合，显著提升复杂环境下的检测鲁棒性。 ### Key features: - **双流处理架构**：支持 RGB 与 IR 图像的端到端联合处理。 - **兼容 YOLOv8 API**：保留原生接口规范，确保用户迁移成本最小化。 - **多层次融合策略**：提供多种融合模块，包括数据级、决策级、早期特征融合、中期特征融合、极简融合和 DEYOLO、DEFA 等动态增强融合注意力。 ### Tech stack: Python ### Why it matters: YOLOFuse 通过引入双流处理架构和多层次融合策略，显著提升了目标检测系统的鲁棒性和适应复杂环境的能力，适用于安防监控、灾害救援、工业巡检等关键场景。其创新性设计和强大的融合能力使其成为多模态目标检测领域的有力工具。

Overview: RSIGame is an autonomous agentic game development framework that continuously improves the quality of generated games through recursive self-improvement, ensuring they are both playable and robust. Key features: - **Recursive Self-Improvement**: RSIGame continuously improves the game by exploring, diagnosing, and improving issues, ensuring the game is both playable and robust. - **Local and Global Loops**: The game development process is divided into a local "explore-diagnose-improve" loop and a global "track-quality-preserve-best" loop, ensuring the game is both playable and robust. - **Experience Internalization**: RSIGame internalizes successful development experience into the generator, improving game quality across various tasks and engines. Tech stack: Python Why it matters: RSIGame is significant because it addresses the challenge of improving generated games beyond a playable version, ensuring they are both robust and consistent. This framework is particularly useful in game development, where the quality of the game is crucial for player satisfaction and the success of the project. Repository: WenyiWU0111/RSIGame GitHub stars: 168 GitHub description: RSIGame: Autonomous agentic game development with recursive self-improvement

### Overview: Notus is a collection of fine-tuned Large Language Models (LLMs) leveraging SFT, DPO, SFT+DPO, and other RLAIF/RLHF techniques. It adheres to a data-first, human-centric approach, making it suitable for various applications and benchmarks. ### Key Features: - Fine-tuned models using SFT, DPO, SFT+DPO, and RLHF techniques. - Data-first, human-centric approach ensuring high-quality results. - Evaluated with Chat (MT-Bench, AlpacaEval) and Academic (Open LLM Leaderboard) benchmarks. - Named after the ancient Greek god Notus, symbolizing the south wind, contrasting Zephyr, the west wind. ### Tech Stack: Python ### Why it Matters: Notus represents a significant advancement in fine-tuning LLMs while maintaining a data-first approach, making it a valuable resource for developers and researchers in the field of AI and language models.

Overview: This project implements a lossless long-term memory system for a personal AI, designed to preserve every conversation detail and timestamp. It stores raw conversation logs in JSONL format and timestamps everything, allowing users to search by time and context. Key features: - Store every line of conversation in full detail. - Timestamp everything, allowing precise time-based searches. - Use a "Temporal Backbone" index to prioritize time over words in search results. - Inject "where we are" context every turn, ensuring identity and context remain intact. Tech stack: Python Why it matters: This system offers a unique approach to long-term memory by preserving every conversation detail and timestamp, ensuring users can retrieve the exact context and history of their interactions with the AI. Repository: aru-labs/lossless-memory GitHub stars: 147 GitHub description: Lossless long-term memory for a personal AI: never summarize, keep every line, put a timestamp on everything.

### Overview: ProntoQA is a synthetic dataset designed to formally analyze the reasoning capabilities of large language models. It consists of question-answering examples with chains-of-thought that describe the reasoning required to answer the questions correctly. The dataset is generated from syntactically simple sentences, making it suitable for formal analysis. ### Key Features: - **Chain-of-Thought Output:** The examples contain chains-of-thought that describe the reasoning process. - **Semantic Parsing:** The sentences are syntactically simple and amenable to semantic parsing, allowing for formal analysis. - **Formal Analysis:** The dataset is used to formally analyze the predicted chain-of-thought from large language models. ### Tech Stack: Python ### Why It Matters: ProntoQA provides a structured way to evaluate and analyze the reasoning abilities of large language models, contributing to the understanding of how these models process and generate answers.

Overview: Bittune is a local inference service harness that deploys, benchmarks, and tunes vLLM/SGLang inference services on your own GPUs. It uses 10 managed tools to audit the environment, deploy services, benchmark, compare results, and generate auditable reports. Key features: - Local inference service for GPU machines - 10 managed tools for auditing and executing tasks - Realistic performance data from Core's real measurements - Deployments and optimizations always occur on your device, verified and auditable - Supports deployment of Qwen to vLLM, benchmarking, and exporting results Tech stack: TypeScript Why it matters: Bittune provides a reliable, auditable way to deploy and optimize inference services on your own GPUs, ensuring that all performance metrics are based on real-world measurements and leaving no room for manipulation or bias. Repository: bitcloud-ai/Bittune GitHub stars: 129 GitHub description: Local-first engineering agent that deploys, benchmarks and tunes vLLM/SGLang inference services on your own GPUs, with auditable run evidence.

Overview: This project provides simple wrappers around Kaldi's online nnet3-chain decoders, making it easier to use Kaldi's speech recognition tools for developers on GNU/Linux systems. Key features: - Simple wrappers for Kaldi's online nnet3-chain decoders - Support for Kaldi's online GMM decoders - Pre-trained models from the Zamia Speech project for quick setup Tech stack: Python 2.7 or 3.5, NumPy, Cython, kaldi-asr Why it matters: This project simplifies the process of integrating Kaldi's speech recognition capabilities into applications, making it accessible for developers without extensive knowledge of the underlying technology.

Overview: Paper-List-DAILY is an automated system that tracks and organizes the latest arXiv papers across various computer vision and AI topics. It fetches and categorizes papers, enriches metadata, and generates formatted Markdown lists for both GitHub README and GitHub Pages. Key features: - Daily auto-update via GitHub Actions - 20+ research topics covering perception core, 3D and motion, foundation models, systems frontier, emerging fields, and science & safety - Smart link enrichment with arXiv PDF, translation, code, and discussion links - Dual output for both README and Jekyll-powered GitHub Pages - Configurable keywords for search filters - Monthly archives for easy historical browsing Tech stack: Python Why it matters: This tool streamlines the process of tracking and organizing the latest research in computer vision and AI, making it easier for developers to stay updated on the latest trends and applications. Repository: isLinXu/paper-list GitHub stars: 168 Last commit: 2026.10.30 CI: GitHub Actions License: Not specified

Overview: Free3D is a project that synthesizes correct novel views from a single image using Free3D, a method that does not require an explicit 3D representation. This project is based on the work of Chuanxia Zheng and Andrea Vedaldi from the University of Oxford. Key features: - Free3D synthesizes novel views from a single image without the need for a 3D representation. - Supports training and testing on different datasets including Objaverse, OmniObject3D, and Google Scanned Objects (GSO). - Provides batch and single image testing functionalities for quantitative and qualitative results. - Includes training scripts for enhancing pose accuracy and smoothing consistency. Tech stack: Python, PyTorch, and other dependencies. Why it matters: Free3D is significant because it offers a practical solution for consistent novel view synthesis from single images, which is particularly useful in applications requiring accurate and consistent 3D reconstruction from 2D images.

Overview: This project provides a comprehensive guide on building a powerful Question-Answering system that leverages Langchain and large language models, including OpenAI's GPT3. The system is designed to accurately answer questions by processing PDF documents and using vector-based databases for efficient search and retrieval. Key features: - Streamlined API for OpenAI API key integration - Efficient use of Qdrant for hosting embeddings and textual documents - Fast and free embedding model using SentenceTransformers - Integration with Langchain for data processing and pipeline creation - React/TypeScript frontend for easy client-side interaction Tech stack: TypeScript Why it matters: This project serves as a practical example for developers looking to combine Langchain and large language models with other frameworks like LlamaIndex, offering significant cost savings by reducing API call costs. The project's modular design and efficient use of vector-based databases make it a valuable resource for anyone interested in creating high-performance Q&A systems.

### Overview: SOVA ASR is a fast speech recognition solution based on the Wav2Letter architecture, designed as a REST API service that can be customized for various needs. ### Key features: - **RESTful API**: Provides a simple and efficient way to interact with the ASR service. - **Customizable Models**: Allows for easy modification of both the code and models to suit specific requirements. - **Docker-Composed Deployment**: Simplifies the setup process by using Docker containers for easy deployment and management. ### Tech stack: Python ### Why it matters: SOVA ASR offers a fast and efficient speech recognition solution that can be easily customized and deployed using Docker, making it suitable for various applications requiring speech recognition capabilities.

Overview: This project is dedicated to fine-tuning large language models (LLMs) using Kubernetes environments. Key features: - Fine-tuning of LLMs on Kubernetes clusters - Utilization of Runbooks for efficient and scalable deployment - Integration with Kubernetes for seamless integration and management Tech stack: Go Why it matters: This project offers a practical solution for developers looking to deploy and fine-tune LLMs on Kubernetes, streamlining the process and ensuring high performance and reliability.

### Overview: This repository contains the source code for the **SIGIR 2023** paper **'Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited'**. It evaluates and compares ID-based and modality-based recommendation models using large-scale datasets for both text-based and vision-based recommendation tasks. ### Key features: - **ID-based Recommender Models**: PixelRec, MicroLens, NineRec, Tenrec, and AntM2C. - **Modality-based Recommender Models**: BERT-based, RoBERTa-based, ResNet-based, Swin-T-based. - **Datasets**: MIND, HM, and Bili datasets. - **Pre-trained Models**: BERT-tiny, BERT-small, BERT-base, RoBERTa-base, OPT-125M, ResNet18, ResNet34, ResNet50, and Swin-T. - **Training**: End-to-end training using PyTorch. ### Tech stack: Python ### Why it matters: This repository provides a comprehensive evaluation of different recommendation models, highlighting the strengths and weaknesses of both ID-based and modality-based approaches. It showcases the importance of considering multiple modalities and large-scale datasets in recommendation systems, which is crucial for improving the accuracy and relevance of recommendations.

Overview: `StableV2V` is an innovative method for video editing that ensures shape consistency, particularly useful in scenarios where user prompts lead to significant shape changes. It demonstrates superior flexibility in various downstream applications and is widely applicable. Key features: - Shape Consistency Editing - Sketch-based Editing - Video Inpainting Tech stack: Python Why it matters: `StableV2V` addresses the challenge of maintaining shape consistency in video editing, making it highly valuable in applications where user prompts can cause significant changes to the edited video. Repository: AlonzoLeeeooo/StableV2V Primary language: Python GitHub stars: 168 GitHub description: The official implementation of the paper titled "StableV2V: Stablizing Shape Consistency in Video-to-Video Editing". Repository: AlonzoLeeeooo/StableV2V Primary language: Python GitHub stars: 168

### Overview: This project aims to integrate AI-based features into the [Monika After Story mod](https://github.com/Monika-After-Story/MonikaModDev) using the submod API. It leverages multiple AI models for various functionalities such as text generation, TTS, speech recognition, and emotion detection. ### Key features: - **Text Generation**: Uses `text-generation-webui` for generating text. - **Text-to-Speech (TTS)**: Utilizes `TTS Coqui-AI` and `Tortoise-TTS` for converting text to speech. - **Speech Recognition**: Employs `OpenAI Whisper` with a microphone option. - **Emotion Detection**: Integrates an emotion detection model from `michellejieli/emotion_text_classifier`. - **NLI Classification**: Uses `deberta` for natural language inference tasks. ### Tech stack: Python ### Why it matters: This project enhances the mod with advanced AI capabilities, making it more interactive and engaging for players. It offers new ways to interact with the game and Monika, improving the overall experience.

### Overview: Collaborative Gym (Co-Gym) is a framework designed to enable and evaluate human-agent collaboration. Unlike traditional chatbots, Co-Gym allows both human users and AI agents to exchange messages and take actions in a shared workspace. It supports the collaboration process where the human user can intervene in real-time, making it suitable for improving collaborative agents and evaluating human-agent collaboration. ### Key Features: - **Collaborative Environment**: Enables human users and AI agents to collaborate in a shared workspace. - **Bayesian Rating System**: Decouples the contribution of the human and the agent within the same collaborative episode. - **Real User Trajectories Dataset**: A subset of real user trajectories collected through the platform to support research on human-agent collaboration. - **Co-Gym (real)**: Available for travel planning and tabular analysis, open-sourced UI components for developers to build upon. ### Tech Stack: - Python ### Why It Matters: Collaborative Gym provides a platform for evaluating and improving human-agent collaboration, making it valuable for researchers and developers looking to enhance the performance of collaborative agents.

### Overview: Infinite Pokémon is a Pokémon-style adventure built with Codex, featuring AI-generated maps, NPCs, branching saves, and multiplayer. The game never ends, offering a continuous and evolving narrative through play. ### Key Features: - **AI-Generated Maps and NPCs**: Generates unique environments and characters based on your gameplay. - **Branching Saves**: Allows you to save your progress and continue from any point. - **Multiplayer**: Connects multiple players to share and explore the same world. - **Local AI**: Continuously generates new content based on your interactions. ### Tech Stack: TypeScript ### Why It Matters: Infinite Pokémon is an experimental fan project that explores the idea of a game world that keeps growing as you play. It combines the simplicity of Pokémon-style gameplay with the potential for continuous story generation, making it a unique and engaging experience.