Projects

### Overview: This project is a comprehensive collection of papers and resources related to Large Language Models (LLMs) for Automated Program Repair (APR). It serves as a curated summary and update system for new research in the field, ensuring quality through human curation. The project aims to provide a systematic review of the latest advancements in LLMs for program repair, including various repair scenarios and techniques. ### Key Features: - **Semantic Bug**: Correcting logical errors in the program's logic. - **Security Vulnerability**: Fixing vulnerabilities that could lead to security breaches. - **Syntax Error**: Correcting syntax-related issues. - **Programming Problem**: Addressing problems related to programming logic and structure. - **Static Warning**: Fixing warnings that do not affect the program's functionality. - **Self-Debug**: Using the model to debug and fix errors in the code. - **Type Error**: Correcting type-related issues. - **Web UI Test**: Fixing issues related to web user interface tests. - **Repository-level Issue**: Addressing issues specific to the repository. - **Smart Contract**: Fixing smart contract-related errors. - **Hardware Bug**: Correcting hardware-related issues. - **Performance Bug**: Fixing performance-related issues. - **API Misuse**: Correcting issues related to API misuse. - **Formal Specification**: Fixing formal specification-related errors. - **Crash Bug**: Fixing crash-related issues. - **Error-handling Bug**: Correcting error-handling issues. - **Formal Proof**: Ensuring formal proofs are correct. - **Translation

Overview: This repository provides a conceptual guide and practical resources for designing robust multi-agent systems, focusing on the challenges and effectiveness of orchestrating, governing, and scaling systems with multiple specialized agents. Key features: - Conceptual guide for multi-agent system design - Practical resources for building and deploying agents - Guidance on designing for change and balancing extensibility with pragmatic engineering - Grounded in production-scale, real-world solutions Tech stack: Python Why it matters: This guide offers actionable advice for architects, engineers, and data scientists familiar with agentic services, aiming to extend or build multi-agent architectures from the ground up. Repository: microsoft/multi-agent-reference-architecture GitHub stars: 248 GitHub description: Guide for designing adaptive, scalable, and secure enterprise multi-agent systems Primary language: Python GitHub description: Guide for designing adaptive, scalable, and secure enterprise multi-agent systems

### Overview: Vector Storage is a lightweight and efficient vector database that stores document vectors in the browser's IndexedDB. It leverages OpenAI embeddings to convert text documents into vectors and enables semantic similarity searches based on cosine similarity. ### Key features: - **Document Vector Storage**: Stores and manages document vectors in IndexedDB. - **Semantic Similarity Searches**: Utilizes cosine similarity to find similar documents based on their content. - **Automatic Storage Management**: Automatically removes least recently used documents when storage exceeds the specified limit. - **OpenAI Embeddings**: Converts text documents into vectors using OpenAI models for accurate and relevant search results. ### Tech stack: TypeScript ### Why it matters: Vector Storage simplifies the process of performing semantic similarity searches on text documents, making it easier to find relevant information quickly and efficiently. The use of OpenAI embeddings ensures that the search results are contextually relevant, enhancing the user experience.

### Overview: FreeFuse is a training-free framework designed for multi-subject text-to-image generation. It enables seamless fusion of multiple LoRAs without additional training or user-defined masks, achieving superior identity preservation and compositional fidelity compared to existing methods. ### Key Features: - **Adaptive Token-Level Routing (FreeFuseAttn)**: Dynamically matches subject-specific tokens to their corresponding spatial regions during inference, bypassing the need for external segmentors. - **Training-Free**: No additional training, model modifications, or user-defined masks required. - **Compatibility**: Supports ComfyUI, Z-Image-turbo in diffusers, and FLUX.2-klein-4B/9B in both diffusers and ComfyUI. - **Experimental MOSS Image Base/SFT Support**: Experimental support for MOSS Image Base/SFT in diffusers and ComfyUI. - **Anima Base v1.0 Support**: Anima Base v1.0 support in ComfyUI. - **ControlNet+FreeFuse Integration**: Seamless integration of ControlNet+FreeFuse in ComfyUI. - **Native MiniMax H3 Text-to-Audio-Video Support**: Native support for MiniMax H3 in ComfyUI and diffusers. ### Tech Stack: - Python ### Why It Matters: FreeFuse simplifies multi-subject text-to-image generation by enabling seamless fusion of multiple LoRAs without the need for additional training or user-defined masks, making it highly practical and efficient.

Overview: This project aims to provide a comprehensive guide to understanding and implementing the RAG (Retrieval-Augmented Generation) architecture, focusing on the key components and techniques involved in building an effective RAG system. Key features: - Query Transformations - Intelligent Routing - Indexing - Retrieval and Re-ranking - Self-Correcting Agentic Flows - End-to-End Evaluation Tech stack: Jupyter Notebook Why it matters: This project serves as a practical resource for developers and researchers interested in creating and optimizing RAG systems, providing a clear and detailed overview of the essential components and techniques needed to build a robust RAG-based AI system. Repository: FareedKhan-dev/rag-ecosystem GitHub stars: 248 GitHub description: Understand and code every important component of RAG architecture

Overview: Free4D is a novel tuning-free framework designed for 4D scene generation from a single image, addressing the limitations of existing methods by leveraging pre-trained foundation models for efficient and generalizable 4D scene representation. Key features: - **Image-to-Video Diffusion Models**: Animating the input image to generate a video sequence. - **Spatial-Temporal Consistency Initialization**: Using a point-guided denoising strategy for spatial consistency and a novel latent replacement strategy for temporal coherence. - **Consistent 4D Representation**: Proposing a modulation-based refinement method to mitigate inconsistencies and fully leverage generated information. Tech stack: The project primarily uses Python, with additional dependencies like torch for deep learning. Why it matters: Free4D offers a significant advancement in single-image-based 4D scene generation, providing efficient and generalizable 4D scene representation, which is crucial for applications requiring real-time and controllable spatial-temporal rendering.

Overview: Kaldi-ONNX is a tool designed to convert Kaldi speech recognition models to ONNX models, enabling faster inference on various devices. Key features: - Converts Nnet2 and Nnet3 models - Supports almost all components in Nnet2 and Nnet3 models - Optimizes for Android, iOS, Linux, and Windows devices with highly optimized NEON kernels - Converts RNN and TDNN networks to DAG-like networks Tech stack: Python Why it matters: This tool simplifies the porting of Kaldi models to ONNX, making speech recognition more portable and efficient, especially on heterogeneous devices. Repository: XiaoMi/kaldi-onnx GitHub stars: 248

Overview: Langfuse Docs is an open-source platform that evaluates, traces, and improves LLM applications using a unified open platform. Key features: - Markdown URL endpoints (`.md` suffix) - Local development setup - Formatting with Prettier - Python cookbooks for Jupyter notebooks - Media assets stored in public/images/ and hosted on Cloudflare R2 Tech stack: Fumadocs, Next.js App Router, Shadcn UI, Tailwind CSS Why it matters: Langfuse Docs provides a comprehensive solution for developers working with large language models, offering a unified platform for evaluation, tracing, and improvement of LLM applications. Repository: langfuse/langfuse-docs GitHub stars: 247 GitHub description: 🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.

### Overview: ShapeLLM is the first 3D Multimodal Large Language Model designed for embodied interaction, focusing on universal 3D object understanding with 3D point clouds and languages. ### Key Features: - **Single-View Colored Point Cloud Input:** ShapeLLM supports inputting 3D point clouds from RGBD cameras, making it easy to obtain. - **Robust 3D QA Benchmark:** A novel 3D Multi-Modal (3D MM) benchmark for evaluating 3D object understanding. ### Tech Stack: - Python ### Why It Matters: ShapeLLM is a groundbreaking model that addresses the need for universal 3D object understanding in embodied interaction, significantly advancing the field of multimodal language models.

### Overview: This repository contains a comprehensive review of benchmarks for large language models (LLMs) in software engineering tasks. It includes various categories such as repository and agentic software engineering, program repair, testing and debugging, security, reliability, code understanding, performance optimization, frontend development, and code generation. The benchmarks are categorized by granularity, interaction pattern, evaluation method, and environment, ensuring a thorough review of LLMs' capabilities across different software engineering domains. ### Key Features: - **Repository & Agentic Software Engineering**: - **ProjDevBench** (2026): Measures AI coding agents' performance in end-to-end project development. - **HWE-Bench** (2026): Evaluates LLM agents' effectiveness in hardware bug repair tasks. - **SpecBench** (2026): Assesses LLMs' ability to comply with specifications in complex systems. - **IDE-Bench** (2026): Evaluates LLMs as IDE agents in real-world software engineering tasks. - **RepoGenesis** (2026): Benchmarks the generation of microservices from README to repository. - **SWE-Refactor** (2026): Reviews the refactoring process of code for better understanding and optimization. ### Tech Stack: - **Python**, **JavaScript**, **Java**, **C++**, **Rust**, **Go**, **C#**, **Ruby**, **TypeScript**, **Kotlin**, **Swift**, **R**, **SQL**, **No

### Overview: UniMMAD: Unified Multi-Modal and Multi-Class Anomaly Detection via MoE-Driven Feature Decompression is a state-of-the-art framework for anomaly detection that leverages a mixture-of-experts (MoE)-driven feature decompression mechanism. This unified approach unifies multi-modal and multi-class anomaly detection tasks within a single framework, achieving state-of-the-art performance across various datasets. ### Key Features: - **Unified Multi-Modal and Multi-Class Anomaly Detection**: UniMMAD integrates multi-modal and multi-class anomaly detection tasks into a single unified framework. - **MoE-Driven Feature Decompression**: The core mechanism decompresses multi-modal inputs into compact, general-purpose features, enabling adaptive and disentangled reconstruction tailored to specific domains. - **Efficient Parameter Usage**: A grouped dynamic filtering mechanism and MoE-in-MoE structure reduce parameter usage by 75%, while maintaining sparse activation and fast inference. - **State-of-the-Art Performance**: Achieves state-of-the-art performance on 9 anomaly detection datasets across 3 fields, 12 modalities, and 66 classes. ### Tech Stack: - Python ### Why It Matters: - UniMMAD provides a unified solution for multi-modal and multi-class anomaly detection, simplifying the development process and improving efficiency. - The MoE-driven feature decompression mechanism ensures adaptive and disentangled reconstruction, leading to better performance and faster inference times. Repository: yuanzhao-CVLAB/UniMMAD Primary Language: Python GitHub Stars: 248

### Overview: This repository collects research papers, implementations, and resources related to Foundation Models for Scenario Generation and Analysis in autonomous driving. The repository is maintained by the TUM-AVS (Professorship of Autonomous Vehicle Systems at Technical University of Munich) and will be continuously updated to track the latest work in the community. ### Key features: - **Research Papers**: Collects and categorizes papers related to scenario generation and scenario analysis. - **Implementations**: Includes implementations of Foundation Models used in scenario generation and analysis. - **Datasets**: Repositories of datasets used in scenario generation and analysis. - **Simulators**: Simulators used for testing and evaluating scenario generation and analysis models. - **Benchmark Challenges**: Challenges used to evaluate the performance of scenario generation and analysis models. - **Other Related Topics**: Includes papers on the implementation of Foundation Models, other related topics, and useful resources. ### Tech stack: Python. ### Why it matters: This repository serves as a comprehensive resource for researchers, developers, and practitioners interested in the latest advancements in Foundation Models for Scenario Generation and Analysis in autonomous driving. It provides a platform to stay updated with the latest research, implementations, and resources in the field, facilitating collaboration and innovation.

### Overview: The Global Agent Hackathon is a month-long, open-source AI Agent Hackathon that welcomes builders and dreamers from around the world. The event focuses on agents, RAG, tool use, and multi-agent systems, powered by various AI tools. Participants can win up to $25,000 in cash and credits for their projects, with prizes distributed based on the use of specific tools. ### Key Features: - **Agno**: Showcase projects using Agno. - **Exa**: Free $50 API credits for all participants. - **Firecrawl**: 5k Firecrawl credits for the best project. - **Mem0**: Free 3 months of Mem0 pro plan. - **Lutra AI**: Free 3 months of Lutra AI pro plan. - **Potpie**: $200 Groq API credits for the best project. - **Groq**: $10 Groq API credits for all participants, $300 for the best project. - **MCP Run**: Free 3 month pro account for any participant. - **Graphlit**: Details coming soon. - **Lutra AI**: Details coming soon. - **Groq**: $10 Groq API credits for all participants, $300 for the best project. - **Special Recognitions**: Top 5 projects featured in the Awesome LLM Apps repo, $300 Groq credits, and a feature on the Groq Applications Showcase. ### Tech Stack: The hackathon uses various AI tools and platforms, including Agno,

Overview: DiffSinger is a PyTorch implementation of the DiffSinger: Singing Voice Synthesis via Shallow Diffusion Mechanism (focused on DiffSpeech) research paper. This project focuses on improving the efficiency and performance of singing voice synthesis. Key features: - Shallow version of DiffSpeech (Shallow Diffusion Mechanism) - Auxiliary decoder from FastSpeech2 - Easier trick for boundary prediction of K - Not yet implemented: Multi-speaker training Tech stack: Python Why it matters: This implementation aims to provide a more efficient and performant solution for singing voice synthesis, potentially improving the quality and speed of synthesized voices compared to existing methods. Repository: keonlee9420/DiffSinger Primary language: Python GitHub stars: 247

Overview: ArcRift is a local memory layer for AI coding tools that syncs conversations and context from browser chats to your local IDE agents, powered by a SQLite knowledge graph. It eliminates the need for cloud storage, subscriptions, or re-explaining yourself. Key features: - Captures AI conversations and builds a searchable knowledge graph. - Automatically injects the right context into every new prompt. - Supports multiple AI tools and coding tools. - No cloud storage, no subscriptions, no re-explaining. Tech stack: TypeScript. Why it matters: It solves the problem of AI forgetting context between sessions, ensuring that your conversations and context are always available and relevant. Repository: Eshaan-Nair/ArcRift Primary language: TypeScript GitHub stars: 246 GitHub description: Persistent local memory layer for AI. ArcRift uses a extension and a native MCP server to sync context and decisions from your browser chats (Claude, ChatGPT, DeepSeek) straight to your local IDE agents (Cursor, Claude Code, Windsurf) powered by a local SQLite knowledge graph.

Overview: This project provides React hooks for in-browser Speech Recognition and Speech Synthesis, making it easy to integrate speech capabilities into your applications. Key features: - `useSpeechSynthesis`: A hook for the browser's SpeechSynthesis API, allowing control over text reading. - `useSpeechRecognition`: A hook for the browser's SpeechRecognition API, enabling real-time listening and response. Tech stack: JavaScript Why it matters: This library simplifies the implementation of speech recognition and synthesis features in React applications, making it accessible and easy to integrate into existing projects.

### Overview MolScore is an automated scoring function designed for the evaluation and benchmarking of de novo molecular design models. It facilitates the scoring of *de novo* compounds in the context of generative *de novo* design using generative models. MolScore supports multiple objectives and can be used to benchmark and evaluate different models and objectives. ### Key Features - **Scoring Functions**: Supports RDKit descriptors, synthesizability, 2D and 3D similarity, QSAR, and docking. - **Objective Definition**: Allows the definition of multi-parameter objectives via JSON files, enabling the proposal of new drug design objectives. - **Benchmarking**: Facilitates benchmarking of objectives and generative models using the MolScoreBenchmark mode. - **Curriculum Learning**: Enables the implementation of a sequence of objectives using the MolScoreCurriculum mode. ### Tech Stack Python ### Why It Matters MolScore simplifies the evaluation and benchmarking of generative *de novo* design models, making it easier to compare and evaluate different models and objectives in the context of drug design. The ability to define and benchmark objectives in a standardized manner enhances the reproducibility and comparability of research in the field of de novo drug design. ### Repository - **Repository**: MorganCThomas/MolScore - **Stars**: 245 - **Description**: An automated scoring function to facilitate and standardize the evaluation of goal-directed generative models for de novo molecular design. ### Installation - **PyPI**: `pip install molscore --upgrade` -

### Overview: ScreenMind is an AI-powered screen memory solution that captures, analyzes, and lets you search and chat your screen history locally using Gemma 4. It is 100% local, private, and doesn't rely on cloud dependencies, making it perfect for privacy-conscious users. ### Key Features: - **📸 Smart Capture**: Content-change detection, not a fixed timer. Captures when your screen actually changes. - **🔬 Gemma 4 Vision Analysis**: Every screenshot analyzed: app detection, activity categorization, mood, scene description, spatial layout regions. - **🔍 Hybrid Search**: Semantic embeddings (MiniLM) + FTS5 keyword search. Find anything by meaning, not just keywords. - **💬 Chat with Memory**: Conversational RAG with follow-up support. Ask "what did Alex say on Discord?" → get the actual message. - **🧠 Model Hub**: In-app model download with variant selection (Q4_0, Q8_0, BF16, etc.). Supports Gemma 4 E2B, E4B, and 12B. Chat and Summary are locked with witty brain animations until the model is ready. - **🎙️ Voice Memos**: Hold `Ctrl+Shift+V` → Gemma 4's native audio encoder transcribes. Screenshot captured alongside, full transcript searchable. Audio always saved, even if transcription fails — re-transcribe any time. - **🎤 Meeting Transcription**: Auto-detects Zoom, Teams, Webex, Slack and Discord apps, plus Google Meet, Teams

### Overview: VisualRWKV is a visual-enhanced version of the RWKV language model, designed to handle various visual tasks. ### Key features: - **Visual Instruction Handling**: Enables RWKV to interpret and respond to visual instructions. - **Image Generation**: Can generate images based on textual descriptions. - **Visual QA (Question Answering)**: Answers questions about visual content. - **Visual Captioning**: Converts visual inputs into captions. ### Tech stack: Python ### Why it matters: VisualRWKV enhances the capabilities of RWKV, making it more versatile for applications that require visual understanding and interaction. Repository: **howard-hou/VisualRWKV** GitHub stars: **246** GitHub description: VisualRWKV is the visual-enhanced version of the RWKV language model, enabling RWKV to handle various visual tasks.

Overview: DiffDock-PP is an implementation of a new approach to rigid-body protein-protein docking that leverages diffusion generative models and confidence models to generate and rank binding poses. This work was presented at ICLR 2023 and is based on the paper "DiffDock-PP: Rigid Protein-Protein Docking with Diffusion Models" by Mohamed Amine Ketata et al. Key features: - Diffusion generative models for generating binding poses - Confidence models for ranking binding poses - PyTorch implementation for efficient training and inference Tech stack: Python Why it matters: This approach offers a novel method for protein-protein docking that combines generative models to generate binding poses and confidence models to evaluate their quality, potentially leading to more accurate and reliable docking results compared to traditional methods. The use of diffusion models and confidence models allows for the generation of diverse and high-quality binding poses, which can be crucial for drug discovery and other applications in computational biology.