Projects

Overview: This project is an AI application that utilizes LLMs and Selenium for resume analysis and LinkedIn scraping, reducing the time spent on job hunting by 30%. Key features: - Resume analysis with LangChain and OpenAI integration - LinkedIn scraping using Selenium - Comprehensive resume summarization and strengths/weakness evaluation - Job title suggestions and perfect job matching - Time-saving resume analysis and LinkedIn scraping Tech stack: Python, NumPy, Pandas, LangChain, Large Language Model (LLM), Retrieval-Augmented Generation (RAG), OpenAI, Selenium, AWS, Hugging Face, Streamlit Why it matters: This application simplifies the job-seeking process by providing valuable insights and recommendations, saving time and effort for candidates.

Overview: VideoCoF is a unified video editing method that achieves remarkable performance across various editing scenarios, including single-shot and multi-shot editing, with only 50,000 training pairs. Key features: - Achieves 16x single-shot editing (512 frames) and 14x multi-shot editing (415 frames) on long videos. - Generalizes well to videos longer than the training set. - Utilizes temporal reasoning to enable efficient editing. Tech stack: Python Why it matters: VideoCoF demonstrates the power of temporal reasoning in video editing, enabling efficient and effective editing with minimal training data, making it a significant contribution to the field of video editing and AI-driven content creation.

### Overview: This repository is a collection of papers and code related to text-to-image generation, focusing on various techniques and models for transforming text into images. The papers are organized by the time of submission to arXiv, and the code links are provided for each model. ### Key features: - **General Text to Image**: Models like GAN-INT-CLS, StackGAN, StackGAN++, AttnGAN, HD-GAN, StoryGAN, MirrorGAN, and DM-GAN. - **Scene Graph/Layout to Image**: Models such as GAWWN, sg2im, and Text2Scene. ### Tech stack: Python, PyTorch, and TensorFlow. ### Why it matters: This collection is valuable for researchers and developers interested in text-to-image generation, providing a comprehensive overview of existing models and their applications.

### Overview: SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning ### Key Features: - **Self-Play Framework**: SPIRAL enables self-play on zero-sum games, where models play against continuously improving versions of themselves. - **Infinite Curriculum**: The framework generates an infinite curriculum of progressively challenging problems as models adapt to stronger opponents. - **Multi-Agent Multi-Turn**: Models play multiple turns with each other, simulating real-world scenarios. - **Reasoning Development**: Through self-play, models develop advanced reasoning strategies, demonstrating improved reasoning capabilities. ### Tech Stack: - Python ### Why It Matters: Self-play in zero-sum games can naturally induce transferable reasoning capabilities, highlighting a promising direction for autonomous reasoning development.

### Overview: This repository contains a collection of useful datasets and resources for Mongolian Natural Language Processing (NLP). It serves as a valuable resource for researchers, developers, and enthusiasts interested in the Mongolian language and its applications in NLP. ### Key Features: - **Datasets:** - **Mongolian TTS Dataset (DATASET):** 8 hours of Mongolian TTS data from Inner Mongolia University, created from the Mongolian Bible. - **Kalmyk TTS Dataset (DATASET):** 2 hours of Mongolian Kalmyk TTS data from the Kalmyk Bible, used in tugstugi/pytorch-dc-tts. - **Kalmyk STT Dataset (DATASET):** 300 hours of Kalmyk synthetic STT data created using a voice conversion model. - **Eduge News Classification Dataset (DATASET):** 75K news articles with 9 categories, used to train the Eduge.mn production news classifier. - **Government Agency Complaint Dataset (DATASET):** 80K complaints with 5 categories, used for sentiment analysis. - **Personal Names Dataset (DATASET):** 220K Mongolian personal names. - **Clan/Family Names Dataset (DATASET):** 90K Mongolian clan/family names. - **Company Names Dataset (DATASET):** 192K Mongolian company names. - **Provinces (Aimags and Sums) Names Dataset (

Overview: SapienAI is a self-hosted academic chatbot and research workspace that unifies the latest models from OpenAI, Anthropic, Google, and local Ollama models into a single, secure interface. It includes features like real-time audio chat, 100% local data storage, academic paper integration, semantic search, and dedicated research spaces for writing and organizing work. Key features: - Multi-Provider Support: Seamlessly switch between the top models from OpenAI, Anthropic, and Google, and local models via Ollama. - Realtime Audio Chat: Have a natural, spoken conversation with the AI. - 100% Local & Private: All messages, files, and images are stored securely on your own device. - Flexible Hosting: Connect directly to APIs or through services like Azure, AWS, and Google Vertex. - Academic Grounding: Reduce hallucinations by backing every response with real academic papers from its integrated search. - Semantic Search: Instantly find information across all your chats and uploaded documents using natural language. - Integrated Editor: Write your paper with as little or as much AI assistance as you need, supporting Markdown, Typst, and LaTeX. - Unified Document Management: Upload project documents, notes, and sources, and the AI can read, summarize, and answer questions about them. - AI-Powered Literature Reviews: Automatically generate and export literature reviews in Word or Excel format based on uploaded sources. Tech stack: Shell Why it matters: SapienAI is a self-hosted academic chatbot and research workspace that provides

Overview: CIDLib is a comprehensive C++ development environment designed for general-purpose applications. Key features: - Integrated development environment (IDE) - Extensive library support - Cross-platform compatibility - Object-oriented programming features Tech stack: C++ Why it matters: CIDLib simplifies the development process by providing a robust, feature-rich environment that enhances productivity and reliability in C++ projects.

Overview: DiffSHEG is a state-of-the-art approach for generating real-time speech-driven holistic 3D expressions and gestures using diffusion models. This project presents a novel method that leverages diffusion techniques to achieve high-quality results in a real-time application. Key features: - Real-time speech-driven generation of 3D expressions and gestures - Utilizes diffusion models for efficient and high-quality output - Supports both BEAT and SHOW datasets for training and testing Tech stack: Python, PyTorch, CUDA, conda, and various libraries for data processing and model training. Why it matters: DiffSHEG offers a significant improvement in real-time performance compared to previous methods, making it highly relevant for applications requiring quick and accurate 3D expression generation, such as virtual assistants, gaming, and more. Repository: JeremyCJM/DiffSHEG Primary language: Python GitHub stars: 209 GitHub description: [CVPR'24] DiffSHEG: A Diffusion-Based Approach for Real-Time Speech-driven Holistic 3D Expression and Gesture Generation (CVPR 2024 Official Repo)

### Overview: Belullama is a comprehensive AI application that bundles Ollama, Open WebUI, and Automatic1111 (Stable Diffusion WebUI) into a single, easy-to-use package. It simplifies the process of creating and managing conversational AI applications and generating images with minimal setup. ### Key features: - **All-in-One AI Platform**: Integrates Ollama, Open WebUI, and Automatic1111 (Stable Diffusion WebUI) in a single package. - **Easy Setup**: Stand-alone version comes with a simple installer script for quick deployment. - **Conversational AI**: Creates and manages chatbots and conversational AI applications with ease. - **Image Generation**: Generates images using Stable Diffusion models through the Automatic1111 WebUI. - **User-Friendly Interface**: Open WebUI provides an intuitive interface for interacting with language models. - **Offline Operation**: Runs entirely offline, ensuring data privacy and security. - **Extensibility**: Customizable and extendable to meet specific requirements. ### Tech stack: Shell ### Why it matters: Belullama simplifies the process of creating and managing conversational AI applications and generating images, making it a valuable tool for developers and creatives alike. Its comprehensive features and ease of use make it a powerful addition to any developer's toolkit.

### Overview: This comprehensive AI toolkit covers all areas of AI, from foundational tools to advanced architectures, making it essential for developers looking to accelerate their work. ### Key Features: - **Machine Learning Frameworks** - **Data Processing & Management** - **Vector Databases** - **Orchestration & Workflow Frameworks** - **Computer Vision** - **Natural Language Processing (NLP)** - **Reinforcement Learning (RL)** - **MLOps** - **PDF Extraction Tools** - **Retrieval-Augmented Generation (RAG)** - **Evaluation & Testing** - **Monitoring & Observability** - **AI Agents** - **Generative AI** - **Deep Learning** - **Advanced LLM Architectures** - **AI Development Assistants** - **Multimodal AI** - **Edge AI & Mobile** - **Audio & Speech Processing** - **Deployment & Containerization** - **DevOps & Infrastructure** ### Tech Stack: Open-source AI tools and resources. ### Why It Matters: This curated collection provides essential resources for developers to accelerate their work, from foundational tools to advanced architectures, making it a go-to toolkit for anyone interested in AI development.

Overview: nano-VectorDB is a simple, easy-to-hack Vector Database implementation that can handle up to 100,000 vectors and return results in 100 milliseconds. It supports multi-tenancy for prototyping and is compatible with Python 3.9.11. Key features: - Single-dependency (`numpy`) - Handles up to 100,000 vectors - Query time: 100 milliseconds - Multi-tenancy support Tech stack: Python Why it matters: nano-VectorDB is suitable for quick prototyping and can handle a large number of vectors efficiently, making it ideal for applications requiring fast vector queries. Repository: gusye1234/nano-vectordb Primary language: Python GitHub stars: 209 GitHub description: A simple, easy-to-hack Vector Database

### Overview: Iris-3B is a 3B-parameter diffusion transformer designed for pixel-space generation and general vision tasks. It generates every pixel directly without the need for a latent space, making it an alternative to traditional vision foundation models like DINOv2. The model is capable of estimating monocular depth and restoring and upscaling images. ### Key features: - **Pixel-Space Generation**: Generates every pixel directly, eliminating the need for a latent space. - **Monocular Depth Estimation**: Fine-tuned for monocular depth estimation. - **Image Restoration and Upscaling**: Restores and upscapes images with the same pixel-space prior. ### Tech stack: Python ### Why it matters: Iris-3B offers a promising alternative to traditional vision foundation models, providing a direct pixel-space generation method that preserves image details and can be fine-tuned for various downstream tasks.

### Overview: This repository provides a comprehensive and up-to-date comparison of AI video generation models, focusing on the best models and APIs for developers. It answers the crucial question: "Which model, via which API, at what price, and how fast?" Every model is mapped to the APIs that serve it, offering indicative pricing, speed, and resolution details. ### Key Features: - **Model Comparison**: Detailed comparison of various AI video generation models. - **API Integration**: Models mapped to specific APIs, including pricing, latency, and performance metrics. - **Pricing and Speed**: Prices per second of output video (verified July 2026) and latency information. - **API Documentation**: Links to the APIs and SDKs for easy integration. ### Tech Stack: - Python SDKs and MCP servers for various video generation models. ### Why It Matters: Developers need a clear and up-to-date guide to choose the best AI video generation model and integrate it into their projects efficiently. This repository answers the critical question of "Which model, via which API, at what price, and how fast?" ensuring developers can make informed decisions quickly. ### Repository: - **Repository Name**: Anil-matcha/awesome-ai-video-models - **Stars**: 207 - **Description**: The most complete, up-to-date comparison of AI video generation models — which model, via which API, at what price, and how fast. ### Related Projects: - **awesome-uncensored-ai-video-models**: Filtering-, access-, and licensing-focused companion catalog for local

### Overview: INTERS, or Instruction Tuning for Large Language Models in Search, is a dataset designed to enhance the performance of large language models (LLMs) in information retrieval (IR) tasks. This repository includes a comprehensive dataset of 20 tasks across three fundamental IR categories: query understanding, document understanding, and query-document relationship understanding. The data is derived from 43 distinct datasets with manually written templates, and it has been successfully fine-tuned for various publicly available LLMs, including LLaMA, Mistral, and Phi. ### Key features: - **20 Tasks**: Covering query understanding, document understanding, and query-document relationship understanding. - **43 Datasets**: Derived from diverse datasets, including 20 tasks each. - **Fine-tuned Models**: Successfully applied to LLaMA, Mistral, and Phi, improving their performance in IR tasks. ### Tech stack: Python ### Why it matters: INTERS provides a valuable dataset and framework for researchers and practitioners to explore the potential of instruction tuning to enhance LLMs' proficiency in IR tasks. By addressing the limitations of prompt-based methods, it opens up new possibilities for leveraging LLMs in information retrieval, potentially leading to more effective and efficient search results.

Overview: Referring Image Matting (RIM) is a new task that aims to extract the accurate alpha matte of a specific object in an image, matching a given natural language description. This paper introduces a large-scale dataset called RefMatte and a method called CLIPMat to serve as a baseline for RIM. Key features: - RefMatte: A large-scale dataset for Referring Image Matting. - CLIPMat: A method to baseline Referring Image Matting. Tech stack: Python, PyTorch, and other open-source tools. Why it matters: The new task of Referring Image Matting and the associated RefMatte dataset and method will open new research directions in the field of image matting and provide a valuable resource for future studies.

### Overview: Verbatim RAG is a provenance-first extractive RAG system that retrieves documents, selects answer-relevant passages, and returns source excerpts with citations. It is built to ensure provenance and is evaluated on the 100-row ACL-Verbatim benchmark, achieving a high micro Word-F1 score of 53.6, surpassing the strongest evaluated LLM extractor. The system supports local ModernBERT or optional LLM-assisted extraction, and can be used with static rendering for CPU execution after model weights are available. ### Key features: - **Provenance-first Extraction**: Retrieves relevant documents and returns evidence text from the source. - **Static Path Extraction**: Built-in verified extraction paths ensure the integrity of the evidence text. - **Contextual Rendering**: Contextual mode generates fixed framing around cited excerpts. - **Citations and Correctness**: Source citations and highlights are returned, but source truth, retrieval recall, relevance, completeness, and entailment are not guaranteed. ### Tech stack: Python ### Why it matters: Verbatim RAG is designed to ensure the integrity of the evidence by returning the text directly from the source, providing a provenance guarantee. It is evaluated on the 100-row ACL-Verbatim benchmark, achieving a high micro Word-F1 score, and is also evaluated on the paper's 150M-parameter model, demonstrating its effectiveness in various benchmarks.

Overview: Local Multimodal AI Chat is a self-hostable multimodal chat application that allows users to interact with text, images, PDFs, and voice through a single interface powered by local models. It features data privacy, modularity, and persistence. Key features: - Local inference with Ollama API - Optional OpenAI API fallback - PDF chat using RAG with Chroma vector database - Image chat via multimodal models - Voice input with Whisper transcription - Persistent session history in SQLite Tech stack: Python, Streamlit, Ollama, OpenAI API, LangChain, Chroma, Whisper, SQLite, Docker Why it matters: This application provides a secure, customizable, and efficient way to interact with AI models locally, enhancing data privacy and modularity. Repository: Leon-Sander/Local-Multimodal-AI-Chat Primary language: Python GitHub stars: 208 GitHub description: Self-hostable multimodal chat with local LLMs (Ollama/OpenAI): PDF RAG, image chat, and Whisper voice, Streamlit + Docker. README excerpt: A self-hostable, multimodal chat application that runs open-source models locally. Chat with text, images, PDFs, and voice in one interface, powered by local models via Ollama (with an optional OpenAI API fallback), retrieval-augmented generation over your own PDFs, and speech-to-text with Whisper.

### Overview: Mangaba AI é um framework de orquestração multi-agente projetado para otimizar processos e comunicação entre agentes. Com funcionalidades avançadas como flows orientados a eventos, memória persistente, RAG, protocolos A2A/MCP, e suporte a múltiplos provedores LLM, Mangaba oferece uma solução robusta e completa para equipes de desenvolvimento. ### Key Features: - **Flows Orientados a Eventos**: Orquestração de tarefas com flows, onde agentes podem ser chamados, executados e monitorados. - **Knowledge Base**: Base de conhecimento separada da memória, permitindo a integração de diferentes tipos de fontes de dados. - **Memória Unificada**: Classe Memory com recall semântico, extração automática de fatos, e gravação assíncrona. - **CLI**: Comandos para criação, execução, teste, treinamento e configuração de agentes. - **Modelos Locais**: Suporte a diversos modelos de linguagem, incluindo Ollama e servidores OpenAI-compatible. - **MCP (Model Context Protocol)**: Consumo de ferramentas de servidores MCP externos. - **Human-in-the-Loop**: Pausa e devolução de trabalho ao agente com notas do revisor. - **Reasoning e Planning**: Planejamento e critério de planej

Overview: AgentGarten is a project that aims to create interactive environments where agents can learn through experience, focusing on both fidelity and realism. The neural renderer adapts the Cosmos3-Nano model to geometry conditions in three stages, including bidirectional rectified-flow training, autoregressive block-causal training, and adversarial forcing. Key features: - Real-time interactive environments - Fidelity and realism in the environment - Neural renderer for generating agent observations - Training recipes and streaming inference for the neural renderer - Adapted Cosmos3-Nano model for geometry conditions Tech stack: Python Why it matters: AgentGarten provides a framework for agents to learn through experience in environments that are both faithful and realistic, using a neural renderer that adapts a pre-trained model to geometry conditions in three stages. This approach aims to create environments that are both engaging and effective for learning.

### Overview: RAG ReAct Agent 是一个基于本地知识库的智能问答系统，采用 Retrieval-Augmented Generation（RAG）技术，结合了检索和生成两个环节，旨在通过多跳推理增强问答的准确性和可信度。系统支持多种文件格式，包括 Word 和 PDF，以及 OCR 加载功能，确保了文档处理的高效性和准确性。 ### Key features: - **检索质量**：采用 Qdrant 语义检索、SQLite BM25、加权 RRF 和可选 Reranker 等技术，确保检索结果的准确性和相关性。 - **Agent 推理**：自研 ReAct 循环模型，支持多工具并行、上下文预算、长期记忆等特性，实现高效且可靠的推理过程。 - **回答可信度**：通过来源卡片、行内引用、数字一致性检查等手段，保证回答的忠实度和可信度。 - **工程可靠性**：包括持久入库任务、重启恢复、Alembic、备份恢复、JWT、SSE 与发布门禁等措施，确保系统的稳定性和安全性。 ### Tech stack: - Python 3.12 - React - MIT 许可证 ### Why it matters: RAG ReAct Agent 是一个高效、准确且可靠的智能问答系统，适用于各种文档处理场景，能够提供高质量的问答支持，帮助用户快速获取所需信息。其强大的多跳推理能力和可信回答机制，使得它在复杂查询和多源信息整合