Projects

Overview: Mantic is a context-aware code search engine that prioritizes relevance over raw speed. After testing across 5 repositories, it demonstrates superior result quality compared to grep/ripgrep, despite some trade-offs in speed for very large codebases. Key features: - **Semantic Reranking (Hybrid Intelligence)**: Combines heuristic speed with neural understanding. Uses local embeddings (`transformers.js`) to find "conceptually relevant" code even without exact keyword matches. - **Code Intelligence**: Deep understanding of your codebase structure using Tree-sitter. - **Go to Definition**: `mantic goto UserService` returns the exact line number across your entire monorepo. - **Find References**: `mantic references handleLogin` finds every usage, respecting `.gitignore`. - **Learned Context (Team Memories)**: Mantic now remembers which files solved previous queries. These patterns are saved locally (`.mantic/search-patterns.json`) and can be committed to git to share knowledge across your team. - **Python Support**: Now includes first-class support for Python imports in the dependency graph. - **Security & Stability**: - Regex DoS protection for user inputs. - Command injection mitigations for VS Code extension. - Safe fallback for non-git directories (scans allow-listed extensions). Tech stack: TypeScript Why it matters: Mantic is essential for AI agents and developers, providing a more efficient and context-aware search engine that prioritizes relevance over speed. Repository: marcoaapfortes/Mantic.sh GitHub stars:

### Overview: VLM Run Hub is a repository of pre-defined Pydantic schemas designed for integrating visual ETL into Vision Language Models (VLMs), making it easier to extract structured data from unstructured visual domains. ### Key Features: - **Pydantic Schemas:** Pre-defined schemas for structured data extraction from images, videos, and documents. - **Integration with VLMs:** Simplifies the integration of visual ETL into workflows with VLMs. - **Real-world Use Cases:** Optimized for practical use cases, ensuring strong-typed, validated outputs. ### Tech Stack: Python ### Why it Matters: By providing pre-defined Pydantic schemas, VLM Run Hub simplifies the process of integrating visual data into VLMs, ensuring strong-typed, validated outputs that are easy to use and integrate with existing systems.

Overview: This project aims to provide a comprehensive WebUI integrated package for various LLMs, allowing users to access and utilize their own customized models and applications without writing any code. Key features: - Integration of LLMs with a WebUI - Support for multiple mainstream LLM APIs and open-source models - Knowledge base, database, role-playing, image generation, LoRA, and full parameter fine-tuning - Data set creation, live2d, and other advanced applications Tech stack: Python Why it matters: This project offers a user-friendly interface for accessing and utilizing LLMs, making it accessible to developers and researchers without requiring extensive coding knowledge.

### Overview: This repository is a curated list of awesome sentiment analysis frameworks, libraries, software, and academic papers. It covers a wide range of sentiment analysis topics, from modern transformer-based libraries to traditional methods, and includes resources like lexicons, datasets, and pretrained models. The latest update covers recent advancements in sentiment analysis, including new models, techniques, and datasets. ### Key Features: - **Libraries**: Modern transformer-based libraries, traditional sentiment analysis frameworks, aspect-based sentiment analysis tools. - **Resources**: Lexicons, datasets, word embeddings, pretrained language models, multimodal sentiment analysis techniques. - **Multimodal Sentiment Analysis**: Overview, recent models, multimodal LLMs, key research findings, applications. - **Multilingual and Cross-lingual Sentiment Analysis**: State-of-the-art models, recent approaches, performance benchmarks, supported languages. - **LLM Techniques**: Prompt engineering, in-context learning, retrieval-augmented generation, parameter-efficient fine-tuning. ### Tech Stack: Python, JavaScript, Java, C++. ### Why it Matters: This repository is essential for developers and researchers interested in sentiment analysis, providing a comprehensive and up-to-date resource for building and improving sentiment analysis models.

Overview: This project presents a novel method called LaneGCN for predicting motion sequences in videos. It leverages graph convolutional networks to learn lane-level representations, enabling accurate motion forecasting. Key features: - LaneGCN: A method for predicting motion sequences in videos. - Graph Convolutional Networks (GCNs): Used to capture lane-level representations. - Accurate Motion Forecasting: The project's primary focus on predicting motion sequences with high accuracy. Tech stack: Python Why it matters: LaneGCN offers a significant improvement in motion forecasting by leveraging graph convolutional networks, making it a valuable tool for applications requiring accurate video sequence predictions.

Overview: ShipFast is a cutting-edge SaaS platform designed to expedite the development and deployment of Generative AI and other AI services, offering a seamless account management, subscription integration, and a comprehensive suite of features that streamline the process of shipping AI projects. Key features: - Account management (Login/Logout/User Profile) - Subscription management (Stripe) - Content management system (Contentful) - Simple OpenAI API integration - Business idea generator (based on keywords) - Chat with different characters - Chat with your documents (PDF, Word, etc.) - AI writing assistant - AI image generator - Image search engine Tech stack: The front-end is built with React, GraphQL, and Tailwind CSS, while the back-end is powered by Python and Django. The infrastructure is managed using AWS CDK. Why it matters: ShipFast simplifies the process of developing and deploying AI services, making it an essential tool for startups looking to leverage the power of Generative AI without the hassle of setting up a full SaaS solution.

Overview: HybVIO is a research-oriented codebase for visual-inertial odometry and an optional SLAM module. It was published for verifiability and reproducibility of results in a paper on real-time visual-inertial odometry. Key features: - Real-time visual-inertial odometry - Optional SLAM module - Supports benchmarks and EuRoC dataset Tech stack: C++ Why it matters: HybVIO is a valuable resource for researchers and developers interested in real-time visual-inertial odometry and SLAM systems. The code can serve as a baseline in VIO and SLAM benchmarks, and it provides a good starting point for understanding and implementing the methods described in the paper. GitHub stars: 551

Overview: Würstchen is an innovative framework designed to efficiently train text-conditional models by compressing the text-conditional stage into a latent space, achieving a 42x compression factor while maintaining faithful image reconstruction. This approach enables faster and more computationally inexpensive training of the text-conditional part. Key features: - Stage A and B for compressing images - Stage C for learning the text-conditional part in the low-dimensional latent space - 42x compression factor - Fast and cheap training of Stage C Tech stack: Jupyter Notebook Why it matters: Würstchen offers a significant improvement in training efficiency and computational cost, making it a valuable tool for researchers and practitioners in the field of text-to-image models. Repository: dome272/Wuerstchen GitHub stars: 552 GitHub description: Official implementation of Würstchen: Efficient Pretraining of Text-to-Image Models README excerpt: Würstchen is a framework for training text-conditional models that compresses the text-conditional stage into a latent space, achieving a 42x compression factor while maintaining faithful image reconstruction. It is fully integrated into the `diffusers` library, making it easy to use for both image reconstruction and text-conditional generation. Training is significantly faster and cheaper compared to other methods, making it ideal for large-scale applications.

### Overview: Edgar Crawler is the only open-source toolkit that downloads raw and unstructured financial filings from EDGAR and converts them into structured JSON files for easy integration into financial NLP experiments. ### Key Features: - **Seamless Downloading**: Retrieves and downloads financial filings from all US publicly-traded companies based on specified filters. - **Structured JSON Parsing**: Extracts and parses specific key item sections from 10-K, 10-Q, and 8-K filings into a structured JSON format. - **Bootstraps Financial NLP Experiments**: Provides a standardized JSON output for easier integration into research and development. ### Tech Stack: Python ### Why It Matters: Edgar Crawler is the only open-source tool that simplifies the process of downloading and parsing SEC filings, making it easier for researchers and developers to integrate financial data into their NLP experiments.

Overview: TypeDB-ML is a library that integrates machine learning capabilities with TypeDB, a graph database. It provides tools for using graph algorithms and machine learning on TypeDB data. Key features: - NetworkX integration for large graph algorithms - PyTorch Geometric (PyG) integration for building Graph Neural Networks (GNNs) - Example for link prediction using PyG - Example usage of Tensorboard for PyG heteroData Tech stack: Python, TypeDB, NetworkX, PyTorch Geometric Why it matters: TypeDB-ML enables the integration of machine learning with graph databases, making it easier to analyze and predict relationships in large datasets.

### Overview: This repository is a curated list of models and resources related to multimodal modeling, focusing on unified, native, and large language models. It covers traditional multimodal techniques, large language models, unified multimodal models, native multimodal models, and closed-source models. ### Key Features: - **Unified Multimodal Models (UMMs):** A curated collection of unified multimodal models for understanding, generation, reasoning, and editing. - **Native Multimodal Models (NMMs):** A curated collection of native multimodal models that integrate perception, reasoning, and generation. - **Large Language Models (MLLMs):** A curated collection of large language models for various applications. - **Closed-Source Models:** Models that are not open-source. ### Tech Stack: JavaScript ### Why It Matters: This repository serves as a comprehensive guide for developers and researchers interested in multimodal modeling, providing a curated list of models and resources across different paradigms. It helps in understanding the latest advancements and staying updated with the latest models and resources in the field.

### Overview: This repository contains a curated list of papers on LLMs-based multimodal generation, focusing on visual (image, video, 3D) and audio modalities. ### Key features: - **Image Generation**: - **LLM-based**: Explore papers using language models for image generation. - **Non-LLM-based**: Papers using non-LLM techniques such as Clip/T5. - **Datasets**: Include datasets for image generation. - **Video Generation**: - **LLM-based**: Papers utilizing language models for video generation. - **Non-LLM-based**: Papers using non-LLM techniques for video generation. - **Video VAE/Tokenizers**: Papers using video variational autoencoders or tokenizers. - **Audio-Video**: Papers combining audio and video elements. - **Benchmarks**: Compare different methods and models. - **Datasets**: Include datasets for video generation. - **3D Generation**: - **LLM-based**: Papers using language models for 3D generation. - **Non-LLM-based**: Papers using non-LLM techniques for 3D generation. - **Datasets**: Include datasets for 3D generation. - **Audio Generation**: - **LLM-based**: Papers utilizing language models for audio generation. - **Non-LLM-based**: Papers using non-LLM techniques for audio generation. - **Datasets**: Include datasets for audio generation. - **Generation with Multiple Modalities**:

Overview: CodeRabbit is an AI-powered code review tool that enhances code quality and accelerates the review process by integrating with popular version control platforms. Key features: - Intelligent feedback on code changes - Official resources including documentation, blog, FAQs, and GitHub repository - Enterprise configuration examples for real-world project setups - Pre-Merge Check Packs for comprehensive code review practices Tech stack: Node.js, TypeScript, JavaScript, and AI-powered code review Why it matters: CodeRabbit simplifies the code review process, making it more efficient and effective for development teams, especially in enterprise environments. Repository: coderabbitai/awesome-coderabbit GitHub stars: 549

Overview: Ethora is a comprehensive SDK monorepo designed for developers to integrate real-time messaging, AI bots, and backend services into their applications quickly. Key features: - React.js chat SDK - React Native chat SDK - Native Android SDK - Native iOS SDK - WordPress plugin - Full web app (React.js) - SDK testing & demo playground - Backend integration examples - MCP server for AI agents and IDEs - CLI setup tool - Self-hosted server installer Tech stack: JavaScript Why it matters: Ethora simplifies the integration of chat, messaging, and conversational AI into various frontend stacks, making it an essential tool for developers building modern applications.

Overview: SparkDiffusion is a video-generation acceleration framework that achieves a 265× speedup over dense multi-step DiT models, thanks to sparse low-rank attention, few-step distillation, and custom high-performance operators. Key features: - Sparse low-rank attention (RoLa) - Few-step distillation (CrossDistill) - Custom high-performance operators - End-to-end pipeline for Wan 2.1 and Wan 2.2 models Tech stack: Python Why it matters: SparkDiffusion provides a significant speedup in video generation, making it highly beneficial for applications requiring real-time or high-performance video processing.

### Overview: PointMamba is a simple state space model designed for point cloud analysis, achieving state-of-the-art performance on various datasets with minimal computational overhead. ### Key features: - **Space-Filling Curves for Tokenization**: Efficiently tokenizes point clouds into a linear space. - **Non-Hierarchical Encoder**: Simplifies the architecture without sacrificing performance. - **Global Modeling Capacity**: Maintains global modeling capabilities while reducing computational complexity. ### Tech Stack: - Python ### Why it matters: PointMamba offers a straightforward yet effective solution for point cloud analysis tasks, significantly improving performance and reducing computational costs compared to traditional methods.

### Overview: Genblaze is an open-source Python SDK designed for orchestrating generative AI media pipelines, including video, audio, and image processing. It integrates with various providers and platforms, ensuring each output has a provenance manifest that can be embedded into media files or persisted to Backblaze B2 or any S3-compatible store. ### Key Features: - **Provenance by Default:** Every run yields a canonical manifest, which is deterministic and embeddable into media files, ensuring tamper-evident storage. - **Unified Pipeline API:** Supports multiple providers and platforms, including OpenAI, Google, Runway, Luma, ElevenLabs, and Stability Audio, with a single `Pipeline` API. - **Storage First-Class:** Uses durable, credential-free URLs for asset storage, compatible with any S3-compatible store like Backblaze B2, AWS S3, Cloudflare R2, and MinIO. - **Fallback Chains and Conformance:** Supports fallback models and provider-contract tests to catch drift before users encounter issues. - **Replayable Runs:** Captures the full provenance of each run, allowing users to reconstruct the run through the manifest or by feeding the canonical parameters back into the Pipeline. ### Tech Stack: - Python 3.11+ ### Why It Matters: Genblaze simplifies the process of building and orchestrating generative AI media workflows, making it easier to swap providers and ensure provenance. It provides a unified API for multiple providers, supports fallback models, and ensures storage is first-class, making it a

### Overview <i>S</i>³Gaussian: Self-Supervised Street Gaussians for Autonomous Driving is an official implementation of a method designed to decompose street scenes into dynamic 3D Gaussians for autonomous driving applications. This approach eliminates the need for external annotations, such as 3D bounding boxes, making it particularly useful in environments where traditional supervision is not available. ### Key Features - **Dynamic Scene Modeling**: <i>S</i>³Gaussian uses 3D Gaussians to represent dynamic street scenes, enabling accurate scene decomposition without the need for additional supervision. - **Multi-Resolution Encoder**: A multi-resolution hexplane-based encoder is employed to encode 4D grid data into feature planes. - **Multi-Head Gaussian Decoder**: A multi-head Gaussian decoder is used to decode the feature planes into deformed 4D Gaussians, effectively reconstructing the scene. ### Tech Stack The codebase is written in Python, utilizing PyTorch for the deep learning framework. ### Why It Matters <i>S</i>³Gaussian offers a novel and effective solution for autonomous driving tasks by providing a robust method for scene decomposition without relying on external supervision, thus enhancing the overall performance and reliability of autonomous vehicles in dynamic urban environments. ### Repository Repository: nnanhuang/S3Gaussian GitHub Stars: 550 GitHub Description: Official Implementation of Self-Supervised Street Gaussians for Autonomous Driving ### Getting Started To get started with <i>S</i><

### Overview: This project showcases a local RAG (Retrieval Augmented Generation) pipeline using Ollama and LangChain, enabling developers to chat with their PDFs locally. ### Key features: - **100% Local Processing**: All processing happens on your machine, ensuring privacy and security. - **Multi-PDF Support**: Upload and query across multiple documents. - **Multi-Query RAG**: Intelligent retrieval with source citations. - **Advanced RAG**: LangChain-powered pipeline with ChromaDB. - **Two Modern UIs**: Next.js and Streamlit interfaces. - **REST API**: FastAPI backend for programmatic access. - **Jupyter Notebooks**: For experimentation and learning. ### Tech stack: TypeScript ### Why it matters: This project provides a powerful and secure way to interact with PDFs locally, enhancing productivity and security in development workflows.

Overview: MicroDrama AI is an AI-powered tool that automates the entire process of turning any idea into a complete short-form drama video. It uses a multi-agent AI pipeline to handle story development, character design, shot-by-shot storyboarding, frame generation, and video production, all using a single MuAPI key. Key features: - 🎭 Idea to Video: Transforms raw ideas into complete micro-dramas. - 📊 Storyboard: Generates a detailed storyboard from the script. - 📸 Frames: Creates frames for each scene. - 🎨 Video: Produces a finished video clip. Tech stack: Python, Next.js, MuAPI. Why it matters: MicroDrama AI streamlines the video creation process, eliminating the need for manual steps and ensuring consistent character design and narrative structure. This tool is particularly valuable for creators looking to produce high-quality, cinematic short-form dramas efficiently. Repository: Anil-matcha/Open-AI-Micro-Drama-Generator Primary language: Python GitHub stars: 547 GitHub description: AI short drama & micro-drama video generator — turns any idea into a complete short-form drama using multi-agent AI pipeline (screenwriter → storyboard → frames → video). Seedance 2 VIP, Kling 3.0 Pro, Veo 3.1, Sora 2.