Projects

### Overview: OpenQodex is an open-source code review tool that runs on your machine before you push your code changes. It uses scanners and a separate AI reviewer to ensure your code is clean and secure before it's committed. The tool is designed to be integrated with your coding agent or terminal, making it accessible and efficient. ### Key Features: - **Code Scanners**: Checks for common issues like security vulnerabilities, code smells, and more. - **Separate AI Reviewer**: Uses Claude Code or Codex to review your code, providing detailed feedback. - **Integration**: Runs on your machine, integrates with your coding agent, and can be started from your terminal. ### Tech Stack: - TypeScript ### Why It Matters: OpenQodex ensures that your code is clean, secure, and follows best practices before it's committed to the repository. This helps prevent bugs, security vulnerabilities, and improves the overall quality of your codebase.

### Overview: FLAME is a PyTorch implementation of the FLAME 3D head model, a lightweight and expressive generic model trained on over 33,000 3D scans. It combines a linear identity shape space with an articulated neck, jaw, and eyeballs, pose-dependent corrective blendshapes, and additional global expression blendshapes. The model is designed for efficient and expressive 3D face representation and is widely used in various applications, including animation, art, and virtual reality. ### Key features: - **FLAME Model Implementation**: A PyTorch-based implementation of the FLAME model. - **Lightweight and Expressive**: Combines a linear identity shape space with an articulated neck, jaw, and eyeballs, pose-dependent corrective blendshapes, and additional global expression blendshapes. - **3D Scans Training**: Trained on over 33,000 3D scans to ensure high accuracy and realism. - **Pose-Dependent Landmarks**: Uses pose-dependent contours for better facial expression representation. - **Model Variations**: Provides different variations of the model to suit various applications. ### Tech stack: Python ### Why it matters: FLAME is widely used in various applications due to its lightweight and expressive nature, making it suitable for real-time applications and efficient representation of 3D faces. Its ability to handle pose-dependent landmarks and blendshapes makes it highly versatile and useful in a wide range of projects. ### Repository: soubhiksanyal/FLAME_PyTorch - **GitHub Stars**:

Overview: DA-CLIP is a model designed to control Vision-Language Models for Universal Image Restoration, achieving a 5th place result in the NTIRE 2024 Restore Any Image Model in the Wild Challenge. Key features: - Supports real-world mixed-degradation images - Provides a posterior sampling technique for better image generation - Pretrained weights available for the Universal IR model and DA-CLIP Tech stack: Python Why it matters: DA-CLIP offers a robust solution for image restoration in challenging real-world scenarios, improving the quality of restored images significantly compared to existing methods.

### Overview: Tree-of-Thought Prompting is a technique that leverages the Tree-of-Thoughts framework to enhance the reasoning abilities of Large Language Models (LLMs), such as ChatGPT. By incorporating a structured approach to problem-solving, it enables LLMs to rectify their errors autonomously and progressively accumulate knowledge, thereby improving their reasoning capabilities. ### Key features: - **Tree-of-Thought Prompting (ToT Prompting)**: A method that borrows ideas from the Tree-of-Thoughts framework to broaden and enhance the Chain-of-Thought prompting concept. - **Autonomous Error Rectification**: LLMs can correct their errors independently while gradually accumulating knowledge. - **Improvement in Complex Questions**: Demonstrated by improving ChatGPT 3.5's reasoning ability to answer complex questions, such as the one about the ball's location. ### Tech stack: - Python ### Why it matters: Tree-of-Thought Prompting allows LLMs to exhibit superior reasoning abilities by rectifying errors autonomously and progressively accumulating knowledge. This technique can significantly enhance the performance of LLMs in answering complex questions, thereby improving their overall reasoning capabilities.

Overview: DeepStack is the world's leading cross-platform AI engine designed for edge devices, with over 10 million installations on Docker Hub. Key features: - Face APIs - Common Objects APIs - Custom Models - Image Enhance - Scene Recognition - SSL Support - API Key support Tech stack: Python Why it matters: DeepStack provides a powerful, pre-built AI engine that can be deployed on various edge devices, including Linux, Mac, Windows, NVIDIA Jetson, and Raspberry Pi, offering real-time and offline AI capabilities. Repository: johnolafenwa/DeepStack GitHub stars: 819

### Overview: This project implements Llama 3+ inference in a single Java file, providing a practical and efficient solution for Llama 3-based models. The project leverages the GGUF format parser and supports various quantization levels, including Q4_0 and Q8_0, to optimize performance and compatibility. ### Key features: - **Single File, No Dependencies**: The code is entirely contained within a single Java file, eliminating the need for external dependencies. - **GGUF Format Parser**: The project supports the GGUF format parser, which is essential for handling the model's input and output. - **Llama 3+ Tokenizer**: Based on the minbpe tokenizer, this project provides a tokenizer optimized for Llama 3+ models. - **Grouped-Query Attention**: The inference process uses Grouped-Query Attention, enhancing the model's performance. - **Support for Llama 3.1 and 3.2**: The project supports both Llama 3.1 and 3.2, with ad-hoc RoPE scaling for 3.1 and tied word embeddings for 3.2. - **Quantization Support**: The project supports various quantization levels, including Q4_0, Q4_1, Q4_K, Q5_K, Q6_K, and Q8_0, for efficient model loading and inference. - **Fast Matrix-Vector Multiplication Routines**: Utilizes Java's Vector API for fast matrix-vector multiplication, enhancing performance. - **CLI with `--chat` and `

Overview: Papermage is a library designed to support NLP and CV research by processing scientific papers. It provides a unified toolkit for various segmentation types, including pages, rows, sentences, and more. Key features: - Segments of text from scientific papers into pages, rows, sentences, and more. - Supports multiple segmentation types for flexibility. - Provides layers of entities for easy access and manipulation. Tech stack: Python Why it matters: Papermage simplifies the process of segmenting and accessing different parts of scientific papers, making it easier for researchers to analyze and manipulate textual information from their research papers.

### Overview: This project implements the sparse attention pattern proposed by the Deepseek team in their "Native Sparse Attention" paper. It is a Python implementation that can be used for various applications, including language modeling. ### Key features: - Sparse attention implementation - Support for different parameters such as dimension, head count, sliding window size, compression block size, and selection block size - Compatibility with PyTorch framework ### Tech stack: Python ### Why it matters: This project provides a practical solution for implementing sparse attention patterns, which can be particularly useful in scenarios where computational efficiency and hardware alignment are critical.

### Overview: Nano Banana Pro (Gemini 3 Pro Image) 是全网最全提示词的集合，包含超过908个案例，涵盖图像生成、多轮编辑、多图像合成、搜索增强、深度思考、多语言支持和角色一致性等功能。通过GPT Image 2.5在线生图，用户可以免费试用，支持1K-4K输出和参考图编辑。 ### Key features: - 🎨 最高4K分辨率输出 - 🔄 多轮编辑功能 - 🖼️ 多图像合成 - 🔍 集成Google搜索 - 🧠 深度思考生成 - 🌍 多语言支持 - 👥 角色一致性 - 🎮 专业级控制 - 🔒 AI透明度 ### Tech stack: HTML ### Why it matters: Nano Banana Pro 提供了从图像生成到多轮编辑的全面解决方案，适用于设计师和创意工作者，帮助他们快速生成高质量的图像作品。其多语言支持和角色一致性功能使其成为全球设计师的首选工具。

### Overview: Picroft is a pre-built enclosure designed for Raspberry Pi 3, 3B+, or 4, featuring a speaker, microphone, and a simple voice interface that allows users to interact with Mycroft directly. ### Key features: - **Enclosure for Raspberry Pi**: A compact enclosure that fits Raspberry Pi 3, 3B+, or 4. - **Integrated Speaker and Microphone**: Built-in speaker and microphone for easy setup. - **Pre-built Micro-SD Image**: Ready-to-use image for burning onto a microSD card. - **Community Tested Hardware**: Available hardware for microphone and speaker. - **Raspberry Pi 3, 3B+, or 4 Compatibility**: Supports older Raspberry Pi versions. - **Power Supply Requirement**: Requires a 2.5 Amp or better power supply. - **MicroSD Card**: 8 GB or larger recommended. ### Tech stack: Shell-based project. ### Why it matters: Picroft simplifies the process of setting up and using Mycroft, providing a straightforward and customizable interface for anyone interested in a voice-controlled environment.

Overview: This repository contains a PyTorch implementation of DeepLabV3, a state-of-the-art model for semantic segmentation. The model has been trained on the Cityscapes dataset, providing high-quality results for urban street scenes. Key features: - PyTorch implementation of DeepLabV3 - Trained on the Cityscapes dataset - Includes a YouTube video demonstrating results Tech stack: Python Why it matters: This model offers advanced semantic segmentation capabilities, making it valuable for applications in autonomous driving, urban planning, and other fields where accurate image understanding is crucial. Repository: fregu856/deeplabv3 GitHub stars: 816 GitHub description: PyTorch implementation of DeepLabV3, trained on the Cityscapes dataset.

Overview: Helix is a self-healing infrastructure designed for autonomous agents, offering 90.3% auto-recovery and instant fixes for agent payments. It wraps functions, diagnoses errors, and remembers previous fixes, significantly reducing costs and improving performance. Key features: - Self-healing runtime for autonomous agents - 6-stage pipeline for error diagnosis and fix - Gene Map for storing and recalling fixes - 100% PCEC (Performance-Centric Error Correction) on 1,083 base mainnet transactions - 100% performance on 5 frontier LLMs tested - 2,140ms warm-up time → 1.1ms repair time, $0.49 → $0.00 per repair Tech stack: TypeScript Why it matters: Helix offers a robust and efficient solution for autonomous agent payments, ensuring zero cost and zero risk of errors, making it ideal for businesses looking to improve their agent performance and reliability. Repository: usehelix/helix Primary language: TypeScript GitHub stars: 816 GitHub description: Self-healing infrastructure for AI agent payments. 90.3% auto-recovery.

### Overview: Codeman is a self-hosted mission control designed for AI coding agents. It allows you to run Claude Code, OpenCode, Codex, Antigravity, Gemini, Pi, Grok, DeepSeek, OMP, and GitHub Copilot inside persistent tmux sessions. The system streams the real terminal to any browser, ensuring agents remain productive even when you're away. It includes features like real-time agent visualizations, auto-resume functionality, and a responsive touch-optimized terminal. ### Key Features: - **One Dashboard, Ten CLIs**: Run Claude Code, OpenCode, Codex, Antigravity, Gemini, Pi, Grok, DeepSeek, OMP, or GitHub Copilot in a single session, locally, in Docker, or over SSH. - **Truly Phone-Friendly**: Touch-optimized terminal with instant local echo, QR login, swipe navigation, and push notifications. - **Runs While You Sleep**: Resumes automatically when a subscription limit resets, ensuring 24/7 unattended runs. - **See Your Agents Think**: Real-time agent visualizations and transcripts for every subagent and teammate. - **Agents Can Drive It Too**: The codeman skill, `codeman agent`, and an MCP server for session-to-session verbs in every CLI mode. ### Tech Stack: - TypeScript - Node.js 22+ - Fastify - TypeScript 5.9 ### Why It Matters: Codeman is essential for developers who need to run multiple AI coding agents simultaneously. It ensures that agents remain productive and responsive

### Overview: Notely Voice is a 100% private AI voice transcription app that converts speech to text across 100+ languages. Built with Compose Multiplatform, it uses Whisper AI for processing, ensuring complete privacy and no cloud uploads. ### Key features: - **Rich Text Editing**: Format your notes with headers, sub-headers, title styling, bold, italic, and underline text, and align text. - **Simple Search**: Find your notes instantly with text search. - **Smart Filtering**: Filter notes by type (Starred, Voice Notes, Recent). - **Organization**: Categorize notes with folders and tags. - **Speech Recognition**: Convert speech to text with high accuracy, works offline, and seamless integration with dictation and audio recording. - **Audio Recording**: Record voice notes and play them back within the app. - **Unlimited Transcriptions**: Transcribe unlimited voice notes to multiple languages. ### Tech stack: Android & iOS ### Why it matters: Perfect for students capturing lectures, professionals documenting meetings, doctors recording patient notes, researchers transcribing interviews, and anyone needing accessible hands-free note-taking across all their devices. The app is completely free, modern, cross-platform, and 100% private, making it a valuable tool for anyone looking to enhance their productivity and accessibility.

### Overview: Enhancr is an elegant and user-friendly GUI for video frame interpolation and video upscaling, utilizing artificial intelligence. It leverages node.js and Electron to provide a seamless experience, powered by state-of-the-art technologies. The app enhances the user experience by speeding up AI processes significantly with NVIDIA's TensorRT and Tencent's NCNN, making it accessible without the need for Docker or WSL. ### Key Features: - Encodes video on the fly and reads frames from source video. - Supports batch processing and live preview. - Allows chaining of interpolation, upscaling, and restoration. - Offers scene detection and customizable color themes. - Features a Discord Rich Presence to track progress and performance. - Supports real-time player with perfect audio, subtitle, font, and attachment support. - Built-in Scene Detection to skip interpolation on scene change frames and mitigate artifacts. - Offers the possibility to load custom ESRGAN models in onnx and pth format and converts them automatically. ### Tech Stack: - Node.js - Electron - NVIDIA's TensorRT - Tencent's NCNN ### Why It Matters: Enhancr provides a stunning experience powered by state-of-the-art technologies without feeling clunky and outdated. It speeds up AI processes significantly with NVIDIA's TensorRT, making it accessible without the need for Docker or WSL. The app enhances the user experience by providing a seamless and customizable interface, ensuring a smooth and efficient video processing experience.

### Overview: BOSS Console is an open-source, multi-platform harness designed specifically for enterprises, science, and research. It provides a native, multi-threaded operator's console for running AI agents like Claude, Codex, Gemini, or OpenCode within a real browser, terminal, editor, and with 200+ MCP tools. The console is built for seamless integration with existing workflows and can be deployed on various platforms including macOS, Windows, and Linux. ### Key Features: - **Real Browser and Terminal Integration**: BOSS Console seamlessly integrates with a real browser and terminal, allowing users to interact with their agents directly from BOSS. - **Extensive Plugin Ecosystem**: The console features a wide range of dynamic plugins that extend its functionality, making it highly customizable and extensible. - **Kotlin and Compose Multiplatform**: BOSS Console is built with Kotlin and Compose Multiplatform, ensuring a smooth and efficient user experience across different platforms. - **Secure and Scalable**: The console is designed to be secure and scalable, making it suitable for both enterprise and research environments. ### Tech Stack: - **Kotlin**: The primary programming language for the console. - **Composable Multiplatform Framework**: Built on top of Kotlin, enabling the creation of a single codebase that can run on multiple platforms. ### Why It Matters: BOSS Console is essential for anyone looking to create and manage always-on agents in a multi-platform environment. Its real browser and terminal integration, combined with a customizable plugin ecosystem, makes it a powerful tool for researchers, scientists, and enterprises.

Overview: TTS Voice Wizard is a versatile tool designed for users of VRChat, allowing them to convert speech to text and vice versa, as well as display song information and other contextual data on their avatar. Key features: - Converts speech to text and vice versa using various methods. - Sends text as OSC messages to VRChat for display on the avatar. - Translates speech from one language to over 50 supported languages. - Offers 100+ different voices with customizable options. - Displays current song on Spotify or browser. - Tracks and displays tracker and controller battery life. - Enables display of heartrate in VRChat's Chatbox with Pulsoid or HRtoVRChat_OSC. - Controls VRChat avatar parameters with voice commands. - Displays customizable and interactive counters for touch events. Tech stack: C# Why it matters: TTS Voice Wizard enhances the VRChat experience by providing accessibility features, allowing users to interact with their avatars in a more engaging way.

Overview: Amazon Bedrock Agents enables you to automate complex workflows and build robust solutions using AI-powered assistants. This repository provides practical examples and best practices for working with Amazon Bedrock Agents. Key features: - Amazon Bedrock multi-agent collaboration - Supervisor mode for complex tasks - Unified conversations with built-in intent classification - Traces for observing agent behavior - Guardrails, security, and privacy Tech stack: Python Why it matters: Amazon Bedrock Agents allows for the creation of AI-powered assistants that can perform complex tasks and interact with various APIs and services, making it a valuable tool for automating workflows and building scalable solutions.

Overview: OpenCode Portal is a web-based interface designed for remote interaction with OpenCode instances, providing session management, real-time chat, file mention support, and more. Key features: - Session management - Real-time chat interface - File mention support (`@filename`) - Model selection - Dark/light theme support Tech stack: React Router, IntentUI, Tailwind CSS, Nitro, OpenCode SDK Why it matters: This project aims to provide a mobile-first, responsive interface for interacting with OpenCode instances, offering a solution to the limitations of the official UI, which is currently under development. Repository: hosenur/portal

### Overview: DS-Harness-Remote is a multi-device remote access solution built on the DeepSeek Harness plugin system. It enables secure, low-latency, and end-to-end encrypted P2P-first network access from PC, Android, and Web clients to Harness and CodeX. The solution supports real-time task management, workspace configuration, and local development services, making it a powerful tool for remote development and automation. ### Key features: - **Remote Access**: Access Harness and CodeX from any device, including PC, Android, and Web. - **Secure Communication**: Uses end-to-end encryption for secure, low-latency communication. - **P2P Network**: Prefers peer-to-peer connections, uses TURN relay nodes, and falls back to server relay when needed. - **Native Interface**: Supports native DeepSeek Harness interface without a separate desktop conversation UI. - **Workspaces and Models**: Manages workspaces and models with the same workflow as local use. - **Terminal Profile**: Runs a terminal-only dsh-TUI profile as a Host. - **Public Listening Port**: No public listening port or public IP required on the Harness Host. - **Workspace Files**: Access workspace files, read-only previews, remote terminals, and authorized local development-service previews. - **CodeX and ACP**: Use the Host's default shell in the session directory and restore the terminal screen on reconnect. - **Android Support**: Keeps per-conversation drafts and reading positions, offers backend-separated search, and local favorites. - **No Public Listening Port**: Connects from anywhere with internet access over a