📦 vg-gui-tasker/vg-gui-tasker

vg-gui-tasker

Teach digital agents to navigate software simply by watching video tutorials. While AI models can answer basic questions about a video, they struggle to translate those visual lessons into actual step-by-step actions on a computer screen. This project closes that gap with a smart search algorithm called TASKER that acts like a human eye, instantly pulling out only the most important video frames needed to complete a task. By filtering out background noise and focusing on key moments, it helps GUI agents click, scroll, and type their way through complex online workflows with incredible accuracy and efficiency.

🆔 @hackernewsgithubprojects

📦 juntaojianggavin/april-medseg

April MedSeg

Mix and match any neural network component for medical image analysis with just a single line of configuration. Setting up medical image segmentation usually means wrestling with rigid, hardcoded architectures, but april-medseg changes the game by turning the entire process into a modular playground. You can instantly swap between over a hundred encoders, dozens of decoders, and specialized skip connections using simple configuration files. Whether you are experimenting with classic architectures, cutting-edge foundation models, or advanced training methods like semi-supervised learning, this tool lets you assemble the perfect pipeline without writing custom glue code. Grab your coffee and start building smarter medical AI systems today.

🆔 @hackernewsgithubprojects

📦 danyalrehman/autobg

autobg

Ditching complex physics simulators to predict how proteins wiggle and fold just got a massive upgrade because we can now generate molecular structures using the same core math that powers modern chat bots. A new project called autobg completely throws out the traditional, slow, math-heavy simulation flows. Instead, it breaks down the exact three-dimensional coordinates of a molecule's atoms and predicts them step-by-step like words in a sentence using a custom transformer. It is incredibly fast, avoids the typical geometric errors that break other models, and its pre-trained model can even predict structures for entirely unseen peptides on the first try.

🆔 @hackernewsgithubprojects

📦 yao-ustc/adapt

ADAPT

adapt is the tool that finally stops vision-language models from hallucinating objects that are not actually in your images. When a visual model writes a long description, its focus naturally drifts away from the picture and starts making things up based on text patterns. This project fixes that by tracking the model's early, highly accurate visual focus to build a stable anchor. If the model's attention starts to wander during writing, the system steps in and gently steers its focus back to the right parts of the image. Give it a spin to build incredibly accurate and visually grounded AI applications.

🆔 @hackernewsgithubprojects

📦 wesammustafa/opencode-everything-you-need-to-know

OpenCode: Everything You Need to Know

The open-source AI coding assistant opencode-everything-you-need-to-know is the ultimate community-maintained guide that finally takes the mystery out of running a terminal-based AI programmer. Think of it as your master blueprint for OpenCode, an open-source terminal app that actually reads your local repository, edits your files, and runs tests directly on your machine. While chat interfaces make you copy-paste snippets, this guide shows you how to safely set up custom commands and automate workflows. My absolute favorite part is how it explains the plan-to-build cycle, letting you design a change risk-free before letting the AI write any code. Check out this guide to turn a chaotic AI assistant into your most...

🆔 @hackernewsgithubprojects

📦 geonhobang/eccv2026_horizon3d

Horizon3D

Detect objects in the far distance with incredible accuracy by combining camera images with radar data. The clever folks at Seoul National University created eccv2026_horizon3d to help self-driving cars see much further down the highway. Typically, merging sparse radar signals with visual feeds gets messy over long distances, but this project uses smart 3D Gaussian shapes to map out the physical space perfectly. It actually smashed the previous best detection benchmarks on a major trucking dataset. While we are waiting on the official code release, it is a brilliant glimpse into how future autonomous vehicles will navigate highways safely.

🆔 @hackernewsgithubprojects

📦 trevtron/rtl-ml

rtl-ml

Identify everyday radio transmissions right from your desk using machine learning on a tiny, low-cost microcomputer. This project uses a basic software-defined radio receiver connected to a Raspberry Pi to listen to the airwaves and classify different types of signals. Instead of heavy, power-hungry deep learning models, it extracts simple mathematical features from raw radio waves and uses a lightweight decision-tree algorithm. This smart approach yields over ninety-six percent accuracy when identifying FM broadcasts, weather alerts, ham radio packets, and wireless sensors, all running locally without any internet connection. It is a brilliant way to explore the invisible electromagnetic spectrum around us.

🆔 @hackernewsgithubprojects

📦 nju-pcalab/orthotryon

orthotryon

orthotryon is the AI fashion generator that tackles multiple styling tasks at the same time without breaking a sweat. Usually, if you want an AI to handle a virtual try-on, reconstruct a garment, and transfer a model pose, you have to run separate, heavy systems that fight each other for resources. This project changes that by using a clever mathematical trick to separate tasks geometrically inside a single, shared model. This keeps the tasks from interfering with each other so you get incredibly clean, realistic clothing swaps and pose changes. It is a massive step forward for lightweight, high-fidelity digital fashion.

🆔 @hackernewsgithubprojects

📦 erwanlorteau/bmad_openclaw

bmad_openclaw

The bmad_openclaw extension is the OpenClaw plugin that finally translates structured agile development into isolated, AI-driven workflows. Instead of letting a single AI chat get cluttered and lose track of your project, this system coordinates specialized agents for every phase of building software. It boots up dedicated personas for analysis, product management, architecture, and coding, passing clean context to each one without any memory bleed. You can run the entire lifecycle from product brief to code review either fully automated or step-by-step with your feedback. It is a brilliant way to keep complex AI development organized and highly precise.

🆔 @hackernewsgithubprojects

📦 ch-bas/cctv-camera-database

CCTV Camera Database

Stop guessing which security camera to buy and search over sixteen hundred different models across dozens of brands in one place. The cctv-camera-database project offers a massive, free, public database packed with detailed technical specifications for home and industrial security cameras. Instead of drowning in marketing hype or getting locked into proprietary ecosystems, you can easily compare real hardware specs like resolution, power options, and local storage capabilities. It is the ultimate open-source resource for smart homeowners and privacy-focused builders to find the perfect camera hardware without the corporate spin.

📰 https://news.ycombinator.com/item?id=48583681

🆔 @hackernewsgithubprojects

📦 nanzhipro/karpathy-llm-wiki-bootstrap-skill

karpathy-llm-wiki-bootstrap-skill

karpathy-llm-wiki-bootstrap-skill is the installable agent tool that compiles your raw articles and research notes into a permanent, self-updating knowledge base. Instead of treating AI as a one-shot chatbot that forgets everything the moment you close the window, this tool lets your editor compile new materials into a structured, interlinked wiki. It automatically organizes everything into dedicated concept, comparison, and synthesis pages while keeping your original sources completely untouched. This means your personal notes compound and grow over time without you having to manually manage broken links and messy formatting. Grab this tool to build a personal encyclopedia that actually remembers for you.

🆔 @hackernewsgithubprojects

📦 keaill/rsiccllm

RSICCLLM

Compare satellite photos taken at different times and automatically describe exactly what changed on the ground. This project is preparing to release a specialized artificial intelligence model that looks at before-and-after aerial images to write natural, detailed sentences about human and environmental shifts, like new buildings appearing or forests shrinking. Instead of forcing humans to manually spot tiny differences across massive landscapes, this system automates the tedious work of tracking geographic updates. Keep an eye on this repository as the creators prepare to share their complete code and dataset following its acceptance at a major computer vision conference.

🆔 @hackernewsgithubprojects

📦 socket7/livediff

Livediff

Watch file changes instantly in your terminal while they happen instead of waiting until after you save. Livediff gives you a minimalist, keyboard-driven interface that updates automatically whenever code generators, style formatters, or search-and-replace scripts modify your files. Unlike heavy desktop programs or basic command-line utilities, it focuses entirely on displaying precise, character-by-character highlights of your current edits with a tiny processing footprint. It is the perfect sidekick for keeping an eye on automated changes in real time. Grab it using your package manager and keep your eyes on your code.

🆔 @hackernewsgithubprojects

📦 andrew-shwetzer/career-ops-plugin

Career Ops Plugin

Career-ops-plugin is the Claude Cowork tool that finally turns your chaotic job search into a structured system. By acting as a private, local AI career assistant, it connects directly with Claude to help you tackle the worst parts of the application process. You simply paste your resume once, and the plugin lets you instantly run deep, honest evaluations of any job posting, auto-generating resumes tailored specifically for applicant tracking systems. It can even scan company portals, track your pipeline, and draft outreach messages. It is a brilliant way to make a stressful job hunt feel organized and automated.

🆔 @hackernewsgithubprojects

📦 mint-sjtu/la4vla

la4vla

Robots usually fail their instructions because they rely on visual shortcuts instead of actually listening to what you tell them to do. The creators of la4vla realized that when a robot gets bombarded with video feeds and movement data simultaneously, the visual noise completely drowns out the spoken instructions. To fix this, they built a smart framework that teaches robots to understand physical tasks by training them on language and movement paths with the cameras completely turned off. By mastering these basic action patterns first, the robot builds a solid understanding of instructions, making its real-world success rate jump by up to forty-five percent once the cameras finally turn on.

🆔 @hackernewsgithubprojects

📦 xgxgnpu/physics-informed-vibe-coding

Physics Informed Vibe Coding

Researchers are now training complex physics-informed neural networks to solve difficult fluid dynamics equations without writing a single line of code by hand. This repository represents the first open-source platform that uses artificial intelligence agents to completely automate scientific machine learning research. Instead of fighting with boilerplate code or manually tuning parameters, a researcher simply brings the core scientific problem, while the AI writes the code, designs the experiments, and refines the models in JAX. It contains GPU-accelerated implementations for advanced neural operators and random feature methods, helping scientists bypass engineering friction to focus entirely on discovery.

🆔 @hackernewsgithubprojects

📦 tsouth89/toolport

Toolport

Connect all your AI tools and clients to a single local gateway so you never have to configure the same servers over and over again. Instead of forcing your AI to load hundreds of tool definitions up front, which quickly burns through thousands of expensive tokens, Toolport uses a clever lazy-discovery trick. It presents a tiny search interface to the AI, allowing it to search and pull in only the exact tools it needs for the task. This cuts down context window overhead by over ninety percent while keeping your sensitive API keys safely locked away in your operating system's secure keychain.

🆔 @hackernewsgithubprojects

📦 aeon-7/vllm-ultimate-dgx-spark

vllm-ultimate-dgx-spark

Nvidia's DGX Spark and Blackwell graphics cards have a massive bottleneck when running heavy artificial intelligence models, but this clever setup completely removes those limits. This project, vllm-ultimate-dgx-spark, packages a custom-built artificial intelligence engine that lets a single container serve a whole fleet of massive language models on next-generation hardware. It patches a nasty bug that used to crash the system when more than thirty-two people connected at once, and introduces a clever memory-saving technique that packs data into four-bit format, effectively tripling your memory capacity. Now your AI assistants stay blazing fast and stable under heavy traffic.

🆔 @hackernewsgithubprojects

📦 zju-real/perceive-to-reason

perceive-to-reason

Visual language models often fail at complex tasks because they try to look and think at the exact same time, mixing up simple observation with deep logic. The perceive-to-reason framework solves this problem by completely separating the visual perception stage from the logical reasoning stage. By using an alternating reinforcement learning strategy, the system trains one model to act as a dedicated perceiver and another as a focused reasoner, passing clean visual details to the logic engine. This division of labor achieves much higher accuracy on complex, high-resolution visual puzzles without getting overwhelmed. It is a brilliant way to build smarter, more analytical visual AI.

🆔 @hackernewsgithubprojects

📦 openrouterteam/awesome-openrouter

awesome-openrouter

Find the perfect, pre-built application to plug your AI models straight into by exploring a hand-curated registry of tools designed for OpenRouter. Instead of building your own user interfaces or coding complex integrations from scratch, awesome-openrouter acts as a master directory connecting you to dozens of projects like terminal coding assistants, browser automation agents, and visual data tools. It solves the massive headache of trying to figure out which applications actually support a unified API key for hundreds of different models. Grab a key, pick any tool from this collection, and start experimenting instantly.

🆔 @hackernewsgithubprojects