Projects

### Overview: Rekal is a command-line tool that captures your AI coding sessions at every git commit, stores them in your repo, and lets any agent recall the reasoning behind a past decision — including the approaches that were tried and rejected. Once that work merges, it travels to your teammates over the same `git push` and `git fetch` you already run — so their agents start from your reasoning, not from nothing. No server, no account, no memory service in the middle. ### Key Features: - **Session Capture**: Rekal captures every AI session at every commit. - **Git Integration**: Stores the sessions in your repo, making them accessible via `git push` and `git fetch`. - **Personalized Memory**: As you recall, the memories stay on your machine only, never synced. - **Adaptive Reasoning**: The reasoning behind every change is recalled in ~7.5K context tokens and a few seconds. - **Team Collaboration**: Team members can recall the conversation behind any change, not just the diff. - **Stop Re-deciding**: Dead-ends already ruled out stay ruled out, preventing re-proposing them. - **Scrubbed Before Storage**: Transcripts pass through secret redaction and home-path anonymization before reaching the database, ensuring privacy. ### Tech Stack: Go ### Why It Matters: Rekal is a game-changer for teams that want to improve their agentic development lifecycle. It ensures that every AI session settles decisions, stops re-deciding, and keeps team memory in git, making it easy to collaborate

### Overview: Mneme 是 DeepSeek Harness 的跨会话记忆插件，用于存储和管理 AI 代理的记忆。它通过自动去重合并、矛盾先冻结等机制确保记忆的准确性，并提供 4-way recall（关键词、向量、BM25、实体）功能，支持本地 SQLite 和 Markdown 镜像，以及一个记忆面板。Mneme 还包括 10 个模型工具、MCP 服务器和命令行工具，旨在提供一个安全、可审计、可回放的 AI 记忆系统。 ### Key features: - **自动去重合并**：后台自动去重合并记忆库中的重复项。 - **矛盾先冻结**：对记忆库中的矛盾信息进行先冻结处理。 - **4-way recall**：支持关键词、向量、BM25、实体四种检索方式。 - **本地 SQLite 镜像**：提供本地 SQLite 镜像存储。 - **Markdown 镜像**：提供可读的 Markdown 镜像存储。 - **记忆面板**：提供一个记忆面板供用户查看和管理记忆。 ### Tech stack: - **JavaScript** ### Why it matters: Mneme 是 DeepSeek Harness 的核心记忆系统，通过提供安全、可审计、可回放的 AI 记忆功能，为 AI 代理提供了强大的记忆管理能力。它确保了记忆的准确性和可追溯性，为 AI 代理的决策提供可靠的数据支持。

### Overview: COSINE is a fine-tuning pre-trained language model approach that utilizes weak supervision to enhance the performance of pre-trained models. This method was presented in the paper "Fine-Tuning Pre-trained Language Model with Weak Supervision: A Contrastive-Regularized Self-Training Approach" at NAACL-HLT 2021. ### Key Features: - **Fine-tuning Pre-trained Language Model with Weak Supervision**: COSINE combines the strengths of pre-trained models with weakly labeled data to improve model performance. - **Contrastive-Regularized Self-Training**: The approach uses a contrastive loss to regularize the self-training process, ensuring better generalization. - **Pre-trained Model**: Utilizes RoBERTa, a pre-trained language model, as the base model. - **Weakly Supervised Data**: The method leverages weakly labeled datasets, such as AGNews, IMDB, Yelp, MIT-R, and TREC, for training. ### Tech Stack: - **Python**: The primary programming language used for implementation. - **PyTorch**: Used for deep learning frameworks. - **Transformers v2.8.0**: The pre-trained model used in the project. ### Why It Matters: COSINE offers a novel way to leverage weakly labeled data, improving the performance of pre-trained models without the need for extensive manual annotation. This approach is particularly useful in scenarios where labeled data is scarce or costly to obtain, making it a valuable tool in natural language processing tasks. ### Repository: The repository is hosted on GitHub at [y

Overview: SpectralQuant is a KV cache compression method for large language model inference that improves upon TurboQuant by focusing on a universal structural property across six models in four architecture families, where key vectors concentrate signal in only 3-4% of the head dimension. This property is identified through a one-time 15-second calibration, allowing for better quality and compression compared to TurboQuant. Key features: - Improved KV cache compression quality - Significant reduction in noise dimensions - Consistent performance across different models and architectures Tech stack: Python Why it matters: SpectralQuant leverages a universal structural property across models to achieve better compression and quality compared to TurboQuant, making it a valuable addition to the field of large language model inference.

### Overview: PosterCopilot is an innovative framework designed for professional graphic design, utilizing Large Multimodal Models (LMMs) to achieve precise layout reasoning and controllable editing. The project is part of the PRLab at Nanjing University and is supported by LibLib.ai and the Chinese Academy of Sciences. It is the official implementation of PosterCopilot, available on GitHub. ### Key Features: - **🎯 Geometrically Accurate Layouts**: Achieves precise spatial positioning through a progressive three-stage training strategy that goes beyond simple regression to distribution-based learning. - **🎨 Aesthetic Reasoning**: Instills human-like design principles and aesthetics through reinforcement learning from aesthetic feedback. - **✂️ Layer-level Control**: Enables precise, fine-grained editing of individual layers while maintaining global visual consistency. - **🔄 Multi-round Iterative Editing**: Supports professional iterative design workflows with multiple refinement rounds on specific elements. - **🎭 Versatile Applications**: Handles complete layout generation, insufficient assets synthesis, theme switching, and canvas reframing. ### Tech Stack: Python ### Why It Matters: PosterCopilot is a cutting-edge solution that leverages LMMs to advance layout reasoning and controllable editing in professional graphic design, making it a valuable tool for designers and developers alike. ### Repository: JiazheWei/PosterCopilot ### GitHub Stars: 204 ### Primary Language: Python ### GitHub Description: Official implementation for PosterCopilot ### Repository: JiazheWei/PosterCopilot ### Primary Language: Python ### GitHub Stars: 204

### Overview: VGGHeads is a large-scale fully synthetic dataset designed for human head detection and 3D mesh estimation. It consists of over 1 million images generated using diffusion models, enabling accurate and efficient detection and reconstruction of heads from single images. ### Key Features: - **Dataset Size:** Over 1 million images generated with diffusion models. - **Head Detection:** Accurate detection of human heads. - **Mesh Estimation:** Reconstruction of 3D head meshes from single images. - **Versatile Usage:** Suitable for various applications in computer vision and AI. ### Tech Stack: - Python ### Why It Matters: VGGHeads provides a comprehensive dataset for researchers and developers to test and improve their models for head detection and 3D mesh estimation. The dataset's large scale and synthetic nature make it a valuable resource for advancing the field of computer vision and AI.

### Overview: Cute 是一个面向全网研究与复杂数据分析的云端智能体，能够从公开网络、YouTube、BiliBili、X、小红书、Hacker News 等多个平台搜集和分析信息，提供实时的、有价值的洞察。 ### Key features: - **全域采集**：调度多个智能体并行采集网页信息、Reddit 社区讨论、X 平台动态、YouTube 字幕内容、小红书互动数据，以及由真实资金与内幕信息定价的 Polymarket 预测赔率。 - **私有文档证据链**：通过多模态 RAG 解析用户私有的 PDF、Word、PPT 与图片，把公域情报与内部知识放进同一条证据链。 - **长期记忆**：沉淀偏好、关键事实与可复用流程，让下一次任务不必从头开始。 - **分析与交付**：结合 E2B 沙箱的代码执行能力与专属数据分析 Agent，完成数据清洗、交叉验证与深度分析，最终以 GenUI 交互画布和 PDF、Word、PPT、HTML 等多格式报告自动交付。 ### Tech stack: JavaScript ### Why it matters: Cute 是一个集成了多种先进技术的云端智能体，能够高效地从多个平台获取和分析信息，帮助用户在复杂的数据环境中快速找到关键情报，从而在决策过程中获得更准确的依据。

Overview: Ragna is a RAG orchestration framework designed to simplify the process of managing and executing RAG (Reactive Analytics Graph) queries. Key features: - RAG orchestration framework - Installation steps - Documentation links - Tutorials: Python API, REST API, Web application - Frequently asked questions - Contribution guidelines - Release notes Tech stack: Python Why it matters: Ragna simplifies the complex process of managing and executing RAG queries, making it easier for developers to integrate and utilize RAG in their applications. Repository: Quansight/ragna GitHub stars: 204 GitHub description: RAG orchestration framework ⛵️

Overview: This repository is a curated list of recent style transfer methods utilizing diffusion models, focusing on image synthesis and video synthesis. Key features: - Foundation & Application Models - Image-Encoder Based - Swap Features In Attention Layer - Optimization-Based Methods - LoRA-Based Methods Tech stack: Unknown Why it matters: The repository provides a comprehensive overview of style transfer techniques using diffusion models, making it valuable for developers looking to enhance their projects with advanced style transfer capabilities.

Overview: Hermes Speech is a plugin that integrates AllModels speech models and voices into the Hermes Agent, providing native TTS, STT, and streaming audio capabilities. Key features: - Native Hermes Agent TTS and STT - Conversational setup for speech - Voice search and previews - Streaming-aware speech - Speech management options Tech stack: Python Why it matters: It allows seamless integration of AllModels speech models and voices into the Hermes Agent, enhancing the agent's ability to provide natural language processing and speech capabilities. Repository: allmodels-io/hermes-speech Primary language: Python GitHub stars: 204 GitHub description: Hermes Agent TTS and STT plugin for AllModels speech models and voices.

### Overview: PersonaMem is a benchmark designed to evaluate language models' ability to infer evolving user profiles and generate personalized responses across various task scenarios. It focuses on persona-oriented, multi-session interactions, utilizing a synthetic dialog generation pipeline to simulate realistic and evolving conversational contexts. ### Key Features: - **Persona-Oriented Interactions**: Users engage with chatbots in multi-session interactions across diverse topics. - **Dynamic Attributes**: Users evolve their preferences over time, and the benchmark assesses how models track and incorporate these changes. - **Multi-Session Benchmark**: The dataset includes annotated questions that evaluate models' ability to adapt to evolving user preferences. ### Tech Stack: Python ### Why It Matters: PersonaMem is crucial for assessing how well language models can personalize responses based on individual user traits and preferences. By focusing on evolving user profiles, it helps in enhancing the effectiveness of LLMs in various applications such as writing assistance, recommendations, and therapy consultations, thereby improving user experience and engagement.

### Overview: Remiqora is a free, open-source local AI music studio that combines the power of ACE-Step, YuE2, and Stable Audio 3 for music generation and production. It features a single interface for all three engines, a built-in multitrack DAW, an AI arranger, and a local language model for prompt generation. ### Key Features: - **One UI**: A unified interface for ACE-Step, YuE2, and Stable Audio 3. - **Mutually-exclusive orchestrator**: Starts one engine and stops another automatically. - **Shared storage**: Tracks are stored in a centralized SQLite database and shared folder. - **DAW on top of generation**: Tracks can be split into stems, edited, and mixed. - **AI arranger**: AI arranges parts based on lane input. - **From an idea**: Generates prompts based on a phrase description. ### Tech Stack: - TypeScript ### Why It Matters: Remiqora simplifies the process of using multiple AI music engines by providing a single, unified interface. It also streamlines the workflow by sharing storage and providing a DAW-like environment, making it easier to manage and edit tracks. The AI arranger feature adds a new dimension to the music production process, allowing for more creative control over the arrangement of tracks.

Overview: This repository contains the TensorFlow 2.0 implementation for the ALBERT model, suitable for fine-tuning tasks. Key features: - ALBERT model pre-training - Domain-specific fine-tuning - Conversion of ALBERT weights from TF-Hub to TF 2.0 Tech stack: Python Why it matters: This repository provides a practical solution for fine-tuning the ALBERT model, making it accessible for various downstream tasks.

### Overview: The AI Learning Kit is a curated collection of AI learning materials designed to provide a structured yet flexible path for AI enthusiasts and professionals. It includes resources for various AI fields such as AI Engineer, ML Engineer, and Data Engineer, along with foundational math and Python courses. ### Key Features: - **Roadmaps:** Structured guides for AI Engineer, ML Engineer, and Data Engineer, covering core concepts and practical applications. - **Mathematics for AI:** Essential math courses from DeepLearning.AI, Khan Academy, and other resources. - **Python for AI:** Specialized courses and resources for beginners and experienced programmers. - **AI Foundations:** Courses from Harvard's CS50 and Udacity's AI Nanodegree, focusing on AI principles and applications. - **Prompt Engineering:** Comprehensive guides and courses from DAIR.AI and OpenAI on prompt engineering for developers. - **Python Libraries for AI:** Resources on the Numpy Stack, Data Science in Python, and Scikit-learn. ### Tech Stack: The AI Learning Kit is built with modern web technologies, including React for the frontend and Node.js for the backend. ### Why It Matters: The AI Learning Kit provides a comprehensive and structured approach to learning AI, covering both foundational knowledge and practical applications. It caters to a wide range of learners, from beginners to professionals, making it an invaluable resource for anyone interested in AI.

Overview: DiffusionART is an efficient, stable, and scalable reinforcement learning algorithm designed for diffusion models. It combines the strengths of forward-process and reverse-process reinforcement learning while mitigating their limitations. Key features: - Efficient training with minimal hyperparameter tuning - Robust to small rollout groups for scalable diffusion reinforcement learning - Supports pairwise feedback and various text-to-image, text-to-video, image-to-video, and text-to-audio/video generation tasks - High performance and stability, requiring minimal computational resources Tech stack: Python Why it matters: DiffusionART simplifies the training process of diffusion models, making them more efficient, stable, and scalable. This algorithm enhances the performance and stability of diffusion models while reducing computational requirements, making it ideal for real-world applications where computational resources are limited.

### Overview: Curated MiniMax H3 video prompts for cinematic, ads, anime, UGC, product videos, and more. Includes playable examples and creator attribution. ### Key features: - Playable examples - Creator attribution - Prompts for various use cases (e.g., Stories & Films, Action & Fantasy, Ads & Products, Music & Performance, Vlog & Social) - Prompts for 550 videos ### Tech stack: JavaScript ### Why it matters: These prompts provide a comprehensive set of video generation ideas for various creative uses, making them a valuable resource for developers looking to enhance their projects with cinematic and engaging video content.

### Overview: IterComp is an innovative framework that leverages the strengths of multiple diffusion models to enhance compositional text-to-image generation. It introduces an iterative feedback learning approach to improve the composition capability of diffusion models. ### Key Features: - **Composition-Aware Model Preferences**: IterComp aggregates preferences from multiple powerful diffusion models to enhance compositional generation. - **Iterative Feedback Learning**: The framework employs an iterative feedback learning method to refine both the base diffusion model and reward models over multiple iterations. - **Gallery of Models**: A curated gallery of six powerful diffusion models is used to evaluate compositional metrics. ### Tech Stack: Python ### Why it Matters: IterComp significantly outperforms existing methods in multi-category object composition and complex semantic alignment, offering new research opportunities in reward feedback learning for diffusion models and compositional generation.

### Overview: ExStruct is an Excel structured extraction engine designed for LLM/RAG pipelines and local automation. It converts Excel workbooks into structured JSON, providing APIs, a CLI for editing, and MCP integration for host-managed integrations. ### Key Features: - **Excel -> structured JSON**: Outputs cells, shapes, charts, SmartArt, table candidates, merged-cell ranges, print areas, and auto page-break areas by sheet or by area. - **Output Modes**: - `light`: Cells + table candidates + print areas + best-effort shapes/charts via direct OOXML for `.xlsx` / `.xlsm`. - `libreoffice`: Best-effort non-COM mode for `.xlsx/.xlsm`. When the LibreOffice runtime is available, it adds merged cells, shapes, connectors, and charts. - `standard`: Excel COM mode with texted shapes + arrows, charts, SmartArt, and merged-cell ranges. - `verbose`: Outputs all shapes with width/height and also emits cell hyperlinks. - **Formula Extraction**: Emits `formulas_map` (formula string -> cell coordinates) via OOXML or the selected openpyxl/COM path. Default enabled in `verbose`. - **Formats**: JSON (compact by default, `--pretty` for formatting), YAML, and TOON (optional dependencies). - **Workbook Editing Interfaces**: CLI for primary edit flows, MCP for host-owned safety controls, and `exstruct.edit` for when needed. ### Tech Stack: Python ### Why It Matters: ExStruct

Overview: This repository is a Humanable Chat Generative-model Fine-tuning tool, designed to facilitate the fine-tuning process for various models like ChatGLM, Qwen, Linly LLaMA, and more. It provides a comprehensive guide on installation, fine-tuning strategies, and inference capabilities. Key features: - Supports multiple fine-tuning strategies including fsdp_zero3, fsdp_zero2, mpdp, mpds, and msds. - Provides detailed documentation on configuration and inference modes. - Includes examples for both single-device and multi-device fine-tuning. - Supports fine-tuning with 8-bit precision for better performance. Tech stack: Python Why it matters: This tool simplifies the complex process of fine-tuning large language models, making it accessible for developers and researchers alike. It offers a robust framework for experimenting with different fine-tuning strategies and optimizing performance on various hardware configurations. Repository: hscspring/hcgf Primary language: Python GitHub stars: 203 GitHub description: Humanable Chat Generative-model Fine-tuning | LLM微调

Overview: MonkeSearch is a fully local, temporally aware natural language file search tool that runs on your PC without requiring a GPU or cloud API calls. It provides fast, accurate results in less than 1 second, making it perfect for AI agents and LLM orchestration pipelines. Key features: - **Local and Temporal Awareness**: MonkeSearch searches files locally and uses temporal context to narrow down results. - **Natural Language Queries**: Users can search for files using natural language, and it converts queries into native Spotlight search predicates. - **Scalability**: The tool is designed to be used offline and can run on potato PCs. Tech stack: Python Why it matters: MonkeSearch simplifies file search by leveraging natural language queries and provides a direct LLM-to-filesystem bridge, making it ideal for AI agents and LLM orchestration pipelines that need to find files autonomously. Repository: monkesearch/monkeSearch GitHub stars: 203 GitHub description: fully local, temporally aware natural language file search on your pc! even without a GPU. find relevant files using natural language in less than 1 second.