Projects

### Overview: DeepSeek Harness (DSH) is a curated ecosystem of plugins, tools, and infrastructure designed for developers. It provides a comprehensive set of features to streamline workflows, enhance productivity, and improve reliability. The ecosystem is built on the official DeepSeek AI runtime, ensuring consistent and reproducible profiles across different environments. ### Key Features: - **Core & Bundles**: Includes a community-maintained reproducible profile installer, the DeepResearch plugin, dual-model plan/execute routing, a zero-dependency tool suite, an adaptive deep-research orchestrator, and a documentation reading mode. - **Agents & Orchestration**: Orchestrates agents and workflows for efficient execution. - **Context & Search**: Enhances search capabilities for better context and results. - **Memory & Knowledge**: Manages and enhances knowledge management. - **Input & Editing**: Provides input editing tools for improved workflow efficiency. - **UI, Themes & Interaction**: Offers a user-friendly interface with customizable themes and interaction options. - **Dashboards & Session UX**: Facilitates real-time monitoring and session-based user experience. - **IDE & Clients**: Supports Integrated Development Environment (IDE) and clients for seamless integration. - **Browser & Remote**: Enables browser-based and remote access to DSH features. - **Models & Inference**: Provides models and inference capabilities for various tasks. - **Git & Engineering**: Integrates with Git for version control and engineering practices. - **Security & Governance**: Ensures secure and governed operations. - **Output & Deliverables**: Manages output and deliverables for project management

Overview: Large Reasoning Models like o1 have demonstrated impressive long-step reasoning capabilities through reinforcement learning. However, they frequently face knowledge insufficiencies, leading to uncertainties and potential errors during prolonged reasoning processes. Key features: - **Search-O1 Framework**: A framework that enhances large reasoning models with agentic search capabilities, enabling them to efficiently find relevant information and integrate it into reasoning processes. - **WebThinker**: A new framework that empowers reasoning models with deep research capability, allowing them to access and utilize a wide range of research papers and resources. - **Search-O1 on AIME24 and AMC23**: The project has been accepted by the EMNLP 2025 conference and is now available on the AIME24 and AMC23 platforms, showcasing its effectiveness in mathematical reasoning tasks. Tech stack: Python Why it matters: The integration of search capabilities with large reasoning models can significantly enhance their performance and reliability, making them more useful and efficient in various applications, particularly in domains requiring long-step reasoning, such as mathematical reasoning and scientific discovery. Repository: RUC-NLPIR/Search-o1 Primary language: Python GitHub stars: 1250 GitHub description: 🔍 Search-o1: Agentic Search-Enhanced Large Reasoning Models [EMNLP 2025] The project aims to bridge the gap between large reasoning models and search capabilities, making them more versatile and effective in real-world applications.

Overview: `repository-harness` is a tool that transforms any software repository into a legible, agent-ready workspace for Claude Code and Codex agents, ensuring that the repository remains the system of record and product documents, decisions, and evidence define the work. Key features: - A compact `AGENTS.md` entrypoint - Repository context and navigation - Optional product, decision, and execution-plan locations - Optional templates for durable plans, decisions, application runbooks, and evidence-backed improvements - An invariant-encoding reference and explicit-only skills for discovery, audit, encoding, and improvement Tech stack: Rust Why it matters: `repository-harness` provides a compact entrypoint, navigable repository map, and optional knowledge-discovery tools to help coding agents solve ordinary engineering problems more effectively. Repository: hoangnb24/repository-harness GitHub stars: 1251

### Overview: Workflow automation has transformed from simple app connections to complex multi-agent systems that reason, plan, and complete tasks autonomously. This curated list includes classic iPaaS, AI-augmented platforms, agentic AI, and more, covering everything from browser automation to AI coding agents and MCP (Model Context Protocol) integration. ### Key Features: - **Classic iPaaS**: Connect SaaS apps with triggers and actions (Zapier, Make) - **AI-augmented automation**: Workflows with LLM steps that reason and adapt (n8n AI agents, Make AI Grid) - **Agentic AI platforms**: Autonomous agents with memory, planning, and tool use (OpenClaw, Lindy) - **Multi-agent systems**: Teams of specialized AI agents collaborating (CrewAI, AutoGen) - **Browser automation agents**: AI that navigates the web like a human (Browser-Use, Skyvern) - **AI coding agents**: Autonomous software engineers (Claude Code, Devin, Cursor) - **MCP ecosystem**: Standardized protocol for connecting AI to any tool or data source ### Tech Stack: - Python, JavaScript, Node.js, and various integrations with popular tools and platforms. ### Why it Matters: Automation today means more than just connecting apps. It means autonomous AI agents that reason, plan, browse the web, write code, and complete complex multi-step work — all without constant human input. This curated list is essential for individuals, developers, and enterprises looking to leverage AI and automation for productivity and efficiency.

### Overview: E-commerce Skills is a comprehensive set of 157 AI agent skills designed to help online sellers across various platforms, including Amazon, Shopify, Etsy, TikTok Shop, and Walmart, achieve significant growth and success. ### Key Features: - **Amazon Review Checker**: Authenticates Amazon reviews for accuracy and authenticity. - **Ebay Review Checker**: Validates eBay reviews for suspicious patterns and buyer manipulation. - **Walmart Review Checker**: Ensures Walmart reviews are genuine and free from suspicious patterns. - **Growth Strategy**: Diagnoses business health and identifies growth opportunities. - **Cross-Border E-Commerce**: Scores international markets and handles VAT/IOSS/GST compliance. - **Marketing Strategy Builder**: Plans omnichannel marketing strategies for paid ads, SEO, email/SMS, content, social media, and influencer partnerships. - **Amazon Profit Calculator**: Provides a complete FBA fee breakdown, including referral fees, fulfillment, storage costs, and break-even analysis. - **PPC Strategy Planner**: Cross-platform PPC strategy for Google, Meta, and TikTok with ROAS targets and budget allocation. - **Product Description Generator**: SEO-optimized product descriptions for any platform. - **Market Expansion**: Plans logistics and localization strategies for international markets. ### Tech Stack: Markdown files that any LLM can read. ### Why It Matters: E-commerce Skills provide a straightforward and efficient way for AI agents to access specialized expertise in e-commerce, making it easier for sellers to improve their business performance and achieve their goals.

### Overview: SincNet is a novel neural architecture designed specifically for processing raw audio samples, offering a compact and efficient method to derive customized filter banks tuned for specific applications. ### Key Features: - **Customizable Filter Bank**: Encourages the first convolutional layer to discover meaningful filters, offering a compact and efficient way to derive a customized filter bank. - **Parametrized Sinc Functions**: Implements band-pass filters using parametrized sinc functions, which are faster and more efficient than standard CNNs. - **Speaker Identification with TIMIT Database**: Provides an example of speaker identification using the TIMIT database, demonstrating the effectiveness of SincNet in real-world applications. ### Tech Stack: - Python 3.6/2.7 - PyTorch 1.0 - PySoundFile - Anaconda Environment ### Why It Matters: SincNet offers a more efficient and compact approach to filter bank derivation, making it particularly useful in applications where real-time processing is crucial. This method significantly reduces computational complexity and memory usage, making it ideal for real-time audio processing tasks. Repository: mravanelli/SincNet GitHub Stars: 1246 GitHub Description: SincNet is a neural architecture for processing raw audio samples. It is a novel Convolutional Neural Network (CNN) that encourages the first convolutional layer to discover more meaningful filters. SincNet is based on parametrized sinc functions, which implement band-pass filters. ### Prerequisites: - Linux - Python 3.6/2.7 - Py

### Overview: Speech Swift is an AI speech toolkit for Apple Silicon, providing on-device speech recognition, synthesis, and understanding capabilities powered by MLX and CoreML. It includes features such as ASR, TTS, speech-to-speech, VAD, and diarization. ### Key features: - **ASR (Speech-to-Text)**: Supports 52 languages, including Qwen3-ASR for 52 languages. - **TTS (Text-to-Speech)**: Native CoreML/MLX transcription with model-generated speaker labels and timestamps. - **Speech-to-Speech**: Enables conversion of speech to text. - **VAD (Voice Activity Detection)**: Helps in identifying active speech in audio. - **Diarization**: Separates conversations into individual speakers. ### Tech stack: Swift ### Why it matters: Speech Swift simplifies the development of AI-driven speech applications, making it easier to integrate speech recognition, synthesis, and understanding into macOS and iOS projects without the need for cloud services or external APIs. This toolkit ensures that speech capabilities are available locally, enhancing privacy and performance.

### Overview: This project provides a minimal set of tools for working with the KITTI dataset in Python, focusing on raw datasets and odometry benchmark datasets. It is designed to facilitate easy access to camera and velodyne data, as well as calibration and timestamps, making it a valuable resource for researchers and developers working with the KITTI dataset. ### Key Features: - Raw datasets and odometry benchmark datasets. - Minimalistic approach, suitable for beginners and advanced users alike. - Supports camera and velodyne data through generators and getter methods. - Assumptions include the availability of calibration data and unchanged directory structure. - Example usage demonstrates how to load and access data from the KITTI dataset. ### Tech Stack: Python ### Why It Matters: The KITTI dataset is widely used in computer vision and robotics research for benchmarking and evaluation purposes. This package simplifies the process of working with this dataset, making it easier for researchers and developers to test and improve their algorithms. ### Repository: - **GitHub Repository:** utiasSTARS/pykitti - **GitHub Stars:** 1246 - **GitHub Description:** Python tools for working with KITTI data ### Installation: - **Using pip:** `pip install pykitti` - **From source:** `git clone https://github.com/utiasSTARS/pykitti.git` and `python setup.py install` ### Assumptions: - Calibration data and timestamps are automatically read. - Directory structure is unchanged from the original KITTI zip files. ### Notation: -

Overview: This repository is a comprehensive toolkit and skill designed for text-only language models, providing image Q&A, long-screenshot OCR, frontend UI restoration, and GUI automation capabilities. It offers optional seamless integration with popular agents like Codex, Claude Code, Pi, Oh My Pi, and OpenCode, enhancing the visual capabilities of text-only models. Key features: - Image Q&A - Long-screenshot OCR - Frontend UI restoration - GUI automation Tech stack: Python Why it matters: This toolkit simplifies the integration of visual capabilities into text-only models, making them more versatile and competitive in the AI landscape. Repository: Anionex/agent-vision-toolkit

### Overview: Audio Webui is a web-based interface for interacting with various audio-related neural networks. It provides a user-friendly interface for managing and utilizing these models. ### Key Features: - Real-time audio processing capabilities - Integration with popular audio models - Easy setup and deployment ### Tech Stack: Python ### Why it Matters: Audio Webui simplifies the process of working with audio-related neural networks, making it accessible to developers and researchers alike. It streamlines the development and deployment of audio applications, enhancing the efficiency and user experience of audio projects. ### Repository: Repository: gitmylo/audio-webui Primary Language: Python GitHub Stars: 1242 ### GitHub Description: A webui for different audio related Neural Networks ### Latest Big Features: - Extensions ### Automatic Installers: [Automatic installers! (Download)](https://github.com/gitmylo/audio-webui/releases/tag/Installers) ### Docker: Community projects: - https://github.com/LajaSoft/audio-webui-docker (Docker compose) - https://github.com/jacen92/audio-webui-docker (Fork of audio-webui) ### Local Install (Manual): #### 🔽 Downloading ```bash git clone https://github.com/gitmylo/audio-webui.git ``` #### 📦 Installing ```bash run.bat ``` #### 🔼 Updating ```bash git pull ``` #### 🏃‍ Running ```bash run.bat ``` ### Google Colab Notebook: Not applicable for

### Overview The ZED SDK is a cross-platform library designed to optimize the performance and accuracy of ZED cameras. It provides tutorials and code samples to help developers quickly integrate ZED SDK into their projects. ### Key Features - **Depth Sensing**: Enhances depth sensing capabilities with improved accuracy and robustness. - **Object Detection**: Boosts object detection performance with better accuracy and reliability. - **Body Tracking**: Improves body tracking accuracy and robustness in crowded scenes and unusual orientations. ### Tech Stack C++ ### Why It Matters The ZED SDK is essential for developers looking to build intelligent robots and spaces that can perceive and interact with the physical world accurately and efficiently. Its features are crucial for applications ranging from robotics to augmented reality and smart home devices. ### GitHub Stars 1,244 ### Repository Repository: stereolabs/zed-sdk ### Primary Language C++ ### GitHub Description The ZED SDK is a cross-platform library designed to get the best out of ZED cameras. In this project, we provide tutorials and code samples to help developers quickly integrate ZED SDK into their projects. ### Release Notes - **ZED SDK 5.5** introduces new body tracking capabilities, improved neural depth precision, and enhanced camera support. - **Camera Support**: Extends support to ZED X One Core, MIPI capture card, and Holoscan sensor bridge. - **Platform Enhancements**: TensorRT 11 support, reduced system package requirements, and faster GPU utilization for GEN_3 positional tracking. - **Fixes

### Overview: Founder OS is an open-source, single-operator business command center designed for solo founders and small teams. It integrates AI-assisted departments like comms, funnel, social, finances, and agents, all managed from a single live dashboard. ### Key Features: - **Unified Dashboard**: A single live dashboard for comms, funnel, social, finances, and agents. - **AI-Assisted Departments**: AI agents cover various business departments, providing a seamless experience. - **Knowledge Graph**: A knowledge core powered by the Optimal Engine, enhancing decision-making. - **Integrated Data**: PostgreSQL + Redis for operational data, Elixir/OTP for Optimal Engine, and G-Brain for knowledge management. ### Tech Stack: Go ### Why it Matters: Founder OS simplifies the complex tasks of running a business, making it easier for solo founders and small teams to manage their operations efficiently. It integrates AI-assisted departments, unified communication, and a comprehensive dashboard, ensuring a streamlined and organized workflow.

### Overview: Crit is a powerful tool for reviewing coding agents' plans, diffs, and running apps in the browser. It simplifies the process by providing a unified interface and adaptive UI for different types of outputs. ### Key features: - **Adaptive UI for each type of output**: Crit offers a distinct review UI tailored to plans, code, and running apps, making the process more intuitive. - **Local-first approach**: Crit runs locally from a single binary, ensuring a seamless experience without needing to deploy changes to a remote server. - **Quickstart**: Easy installation via `brew`, Go, Nix, or downloading the latest release from GitHub. - **Integrations**: Works seamlessly with various agents, including Claude Code, Cursor, GitHub Copilot, OpenCode, OMO, omp, Codex, Gemini, Qwen, Hermes, Windsurf, Cline, Grok, Aider, and Pi. ### Tech stack: Go ### Why it matters: Crit simplifies the review process for coding agents, making it easier for developers to collaborate and improve their work efficiently.

Overview: This repository is a curated collection of awesome resources on image-to-image translation, featuring a diverse range of tools, frameworks, and papers that have been widely recognized for their contributions to the field. Key features: - Generative Adversarial Networks (GANs) - Style Transfer - Image-to-Image Translation Models - Custom Generative AI Image-to-Image Models (joliGEN) Tech stack: Python, PyTorch, TensorFlow, and various open-source frameworks. Why it matters: This collection is essential for developers and researchers looking to explore and implement image-to-image translation techniques, providing a comprehensive overview of the latest advancements and resources in the field.

### Overview: Tango is a text-to-audio (TTA) generation model that utilizes a frozen instruction-tuned language model (LLM) Flan-T5 to generate realistic audio from textual prompts. It achieves state-of-the-art performance across both objective and subjective metrics, demonstrating its effectiveness in generating diverse and high-quality audio outputs. ### Key Features: - **Tango 2**: Built upon the Tango-full-ft checkpoint, Tango 2 underwent alignment training using DPO on the Audio-Alpaca dataset, resulting in improved audio quality and better alignment. - **Tango 2 Model**: Available on Hugging Face, providing pre-trained checkpoints for further research and development. - **Audio-Alpaca**: A pairwise preference dataset used for training Tango 2, enhancing its ability to generate aligned audio. ### Tech Stack: - Python ### Why It Matters: Tango represents a significant advancement in the field of text-to-audio generation, offering high-quality audio outputs from textual prompts. Its ability to generate diverse and realistic audio from a smaller dataset sets it apart from current state-of-the-art models.

Overview: This project aims to demonstrate the real-time emotion recognition of human faces using Python, focusing on mixed emotions and their probabilities. Key features: - Real-time emotion recognition - Mixed emotions detection - Probabilities of detected emotions Tech stack: Python Why it matters: This project showcases the capabilities of emotion recognition technology, particularly in identifying and quantifying mixed emotions on human faces, which has applications in various fields such as psychology, marketing, and security. Repository: otaha178/Emotion-recognition Primary language: Python GitHub stars: 1242

Overview: This project is a Keras implementation of the "One Pixel Attack" technique, which is used to fool deep neural networks by modifying the color of a single pixel. The tutorial focuses on the Cifar10 dataset and uses Differential Evolution to generate adversarial images that reduce the confidence of the neural network's classification. Key features: - Keras implementation - Cifar10 dataset - Differential Evolution algorithm Tech stack: Python Why it matters: This project demonstrates the effectiveness of the One Pixel Attack technique in causing deep neural networks to misclassify images, highlighting the vulnerability of modern machine learning models to adversarial attacks.

### Overview: Supavec is an open-source RAG-as-a-Service platform that allows developers to build powerful vector search and chat applications with any data source at any scale. ### Key features: - **Vector Search Optimisation**: Configurable chunk & overlap, hybrid filter (P95 210 ms), batch embeddings (OpenAI cost -65%) - **Scalable Multi-Tenant Design**: Row-Level Security (RLS) for team-level data isolation, usage-based billing (Free 100 → Basic 750 → Ent 5000 req/mo) - **Real-Time Analytics**: PostHog integration for product & API events tracking (11 schema), request-level tracing, async usage logging - **API Performance**: Streaming or standard responses selectable per request, 100-doc batch embedding reduces latency + cost, background processing keeps critical path lean ### Tech stack: TypeScript, Next.js, Supabase, Tailwind CSS, Bun, Upstash ### Why it matters: Supavec offers a cost-effective, scalable, and easy-to-use alternative to Carbon.ai, enabling developers to build powerful RAG applications with any data source at any scale.

### Overview: LLM-Adapters is a framework that integrates various adapters into large language models (LLMs) to achieve parameter-efficient fine-tuning. It is an extension of the HuggingFace's PEFT library, built on the success of previous work in the field. The project is designed to be user-friendly and accessible, making it easy for developers to fine-tune their models with minimal effort. ### Key Features: - **LLaMa, OPT, BLOOM, and GPT-J Integration**: LLM-Adapters supports a variety of popular LLMs such as LLaMa, OPT, BLOOM, and GPT-J. - **Adapter-Based PEFT Methods**: It leverages adapters to fine-tune models, optimizing performance and efficiency. - **State-of-the-Art Open-Access Models**: The framework includes state-of-the-art models like LLaMa, OPT, BLOOM, and GPT-J, along with widely used adapters such as LoRA, AdapterH, AdapterP, Parallel, Prefix Tuning, P-Tuning, and Prompt Tuning. - **Commonsense170k and LLaMA-13B-Parallel Models**: The project includes a commonsense170k dataset and the LLaMA-13B-Parallel model, which achieves high performance on various benchmarks, including 91% of GPT-3.5 performance on 8 commonsense benchmarks. ### Tech Stack: - Python ### Why It Matters: LLM-Adapters simplifies the process of fine-t

### Overview: KVPress simplifies the development of key-value (KV) cache compression methods for large language models. ### Key features: - Multiple KV cache compression methods implemented. - Benchmarked using 🤗 transformers. - Simplifies the development of new methods for researchers and developers in the field. ### Tech stack: Python ### Why it matters: Simplifies the development of key-value (KV) cache compression methods for large language models, making it easier for researchers and developers to experiment with new techniques.