📦 achimala/jev-paint

Jev Paint turns data uncertainty into digital art

A tiny local web app turns raw statistical uncertainty into actual paintings. It is called Jev Paint. Instead of just showing numbers, it asks an AI model how confident it is about every single pixel in a description. It then turns that hesitation into brush strokes, texture, and color. The less the model is sure, the rougher the paint looks. The more it is sure, the smoother the surface becomes. It runs entirely in your browser with a tiny local helper to handle security rules. You type a prompt, it paints, and you get a unique visual result. It is not a data chart.

🆔 @hackernewsgithubprojects

📦 achimala/jevinci

Jevinci turns math into paintings

Jevinci turns the raw statistical uncertainty of a language model into actual painted artwork. It is a local web app that fetches the probability distributions for every pixel in a grid and renders them using physical painting techniques like impasto lighting and pigment strokes. Instead of showing a dry chart, it uses the model's confidence levels to determine color depth, texture, and relief, creating a visual representation of how the machine thinks. You run a tiny local server to connect your browser to the service, then type a prompt to generate unique, textured pieces that fade in as they render.

🆔 @hackernewsgithubprojects

📦 saadwazir/medcagd

How MedCAGD Maps Eyes with Less Code

MedCAGD is the medical image tool that spots tiny details without the heavy computational cost. It takes standard photos of retinas and pinpoints the intricate network of blood vessels, which is usually a headache for computers because the lines are so thin. The clever trick here is a decoder that gates the data flow, meaning it only looks at pixels that actually matter for the structure. This keeps the model light and fast, proving you can get sharp, accurate segmentation without needing a massive cluster of servers. It is a neat, efficient way to handle complex medical imaging that any developer should study.

🆔 @hackernewsgithubprojects

📦 taeold/djev-run

Djev Run: Playing Snake with an AI on a GPU

Deploy a specialized AI model on Google Cloud Run that answers questions and plays games in your browser. The project takes a large language model and hosts it on a high-end graphics card, turning it into a service that can control classic games like Snake and Tetris. The most striking part is how fast it starts up; by streaming the massive model files directly into memory while the system boots, it cuts the wait time from four minutes down to just forty-seven seconds. You get a server that costs three dollars an hour but scales to zero when idle, so you only pay when you are actually using it.

🆔 @hackernewsgithubprojects

📦 alibaba-damo-academy/clinhallu

ClinHallu: The Truth-Teller for Medical AIs

ClinHallu exposes exactly where medical AI models lie, identifying whether errors stem from misreading images, forgetting facts, or flawed logic. Built by Alibaba DAMO Academy, it tests these models on real medical scenarios using a clever technique called replacement. By swapping in correct data at specific steps, researchers can isolate the source of hallucinations. This matters because a wrong answer in medicine isn’t just a bug; it’s a risk. The project provides the tools to find out if your AI is truly reliable before it touches patient care. If you work with medical AI, you need to know if your model sees clearly, knows its facts, and thinks logically.

🆔 @hackernewsgithubprojects

📦 magiccreator-ai/awesome-gpt-6-astra

The GPT-6 Astra Demo Hub

There is a curated collection called Awesome GPT-6 Astra that gathers real, working demos of the new model in action. It features original projects where creators built games, designed three-dimensional assets in Blender, coded live websites, and produced video content. This is not a list of vague prompts or unverified claims. Instead, you get a verified directory of concrete workflows, complete with links to the original posts, creator credits, and live results if available. The team maintains strict standards to ensure every entry is a genuine, reproducible project rather than a repost.

📰 https://news.ycombinator.com/item?id=49596337

🆔 @hackernewsgithubprojects

📦 lunaaispace-eng/comfyui-degrid

ComfyUI DeGrid

A two-pixel checkerboard pattern is silently baked into every image generated by Qwen and Wan VAEs. ComfyUI DeGrid erases this invisible grid before your upscaler mistakes it for real detail. The node uses a phase-locked detector to find the exact lattice and a notch filter to remove only that specific artifact. It automatically calibrates the strength for each image, ensuring smooth areas stay clean while protecting genuine texture like skin pores or fabric weave. If no grid is detected, the image passes through completely untouched. This simple node fixes a subtle flaw that amplifies during sharpening, giving you cleaner results without manual tuning or extra dependencies.

🆔 @hackernewsgithubprojects

📦 egoist-machines/lodedb

LodeDB: The Tiny Vector Database That Fits in Your Pocket

LodeDB is the embedded vector database that fits your entire search index inside a single pocket. It solves the bloat problem by compressing your data so tightly that a store holding over seventeen thousand documents takes up less space than a single high-resolution photo. You get this compactness without sacrificing speed, thanks to a specialized engine that processes multiple searches at once using your graphics card, making it thousands of times faster than standard tools. It acts as a drop-in replacement for your existing memory layer, meaning you can swap it in with a single command without rewriting your code.

📰 https://news.ycombinator.com/item?id=48668208

🆔 @hackernewsgithubprojects

📦 inclusionai/ming-image

Ming Image: Editable Design Layers from Text

Generate fully editable visual designs directly from text. Ming Image creates complete layouts for interfaces, posters, and infographics, but its standout feature is layer decomposition. It takes a single flat image and splits it into separate, transparent layers. This means you can grab just the text, the background, or a specific graphic element independently. It solves the major problem of editing flat designs by turning static pictures into flexible, multi-layered assets. No need to redraw everything from scratch. You get precise control over every part of the design. It is a powerful tool for anyone who needs to tweak specific elements without starting over.

🆔 @hackernewsgithubprojects

📦 patlux/pi-commandcode-provider

Pi Command Code Provider

You can plug eighteen different AI models, including Claude, GPT, and DeepSeek, directly into your Pi coding agent using a single plugin. Pi Command Code Provider acts as a bridge, letting you pick any model from the Command Code API without managing separate keys or complicated setup. It handles the connection automatically, so you can swap between different thinking styles on the fly. This is perfect for developers who want flexibility without the headache of juggling multiple subscriptions. If you use Pi, this is the easiest way to access top-tier models right in your terminal.

🆔 @hackernewsgithubprojects

📦 mindbox77/zxdesk

A Full Desktop for a 1982 Computer

A full graphical desktop with draggable windows, menus, and a working calendar fits inside forty-eight kilobytes of RAM on a computer from 1982. This project, called ZX Desk, proves that a machine with a three and a half megahertz processor can handle overlapping windows and a file manager. The author wrote this in Z80 assembly to recreate the feel of a classic desktop environment, but without emulators. It runs on real hardware, meaning every single byte of memory is accounted for. The real magic here is the engineering required to make dragging a window smooth on such slow hardware.

📰 https://news.ycombinator.com/item?id=49766676

🆔 @hackernewsgithubprojects

📦 ganyuanran/aegis

Aegis: The Safety Net for AI Coding

Your AI coding agent just stopped guessing at your code and started proving it works. Aegis is a method pack that turns your coding assistant into a disciplined engineer, ensuring it checks your project’s real baseline before making changes and only marking work complete when it has fresh evidence. In rigorous tests, this simple rule reduced unsafe outcomes from thirteen percent to zero. It doesn’t force heavy process on simple tasks, but for complex work, it demands proof before saying done. If you want to stop babysitting your agent and start trusting its results, this is the safety net you need.

🆔 @hackernewsgithubprojects

📦 amal-david/awesome-jev

Awesome Jev: The Curated Hub for Typed AI Decisions

Discover forty-four real-world projects built with Jev, a tool that answers questions about data with yes or no probabilities instead of writing text. This curated collection solves the problem of finding out what people are actually using this unique technology for. It is useful because it links directly to the source code and includes a gallery of demos. You can see Jev choosing browser actions, scoring log entries, or even playing chess against a shared board. The repository helps you understand where a dedicated classifier fits better than a large language model. It is a great starting point for anyone curious about how structured decision-making works in modern applications.

📰 https://news.ycombinator.com/item?id=49802160

🆔 @hackernewsgithubprojects

📦 learnprompt/awesome-seedance

Awesome Seedance: 463 Verified Video Prompts

Four hundred and sixty-three AI video prompts, every single one traced back to the exact post where the creator first shared it, and then stress tested across different models to prove they actually work. That is the core promise of Awesome Seedance, a library built specifically for Seedance video generation. Most prompt collections are just recycled text you find in forums, but this repository only accepts entries where the original media matches the prompt exactly, and if a prompt fails when re run, that failure is logged publicly. You get a verified source rather than a guess.

🆔 @hackernewsgithubprojects

📦 stepfun-ai/step-code

Step Code: The Terminal Agent That Publishes Websites

Hand off a complex coding goal to Step Code and walk away. This open-source tool lives in your terminal, reading your code, writing changes, and running tests all at once. The coolest part? It doesn't just build your app; it publishes your local project as a live, shareable website with a single command. It handles the whole loop from development to delivery right in your command line. If you want an AI assistant that actually ships finished work, not just snippets, Step Code is built for that. It turns your terminal into a complete, autonomous development studio.

📰 https://news.ycombinator.com/item?id=49817387

🆔 @hackernewsgithubprojects

📦 amd-agi/hyperloom

Hyperloom: The AI Tuner for AMD GPUs

Hyperloom acts like an automated performance tuner that watches your AI models run on AMD graphics cards to find and fix slow spots before they become bottlenecks. Instead of you manually tweaking complex settings, the system analyzes the workload and applies optimizations on its own to make everything run faster. It is essentially a tireless assistant that constantly refines how your data flows through the hardware, saving you hours of tedious debugging. If you are running large language models and want them to fly without spending your weekend on manual tuning, this tool does the heavy lifting for you.

📰 https://news.ycombinator.com/item?id=49817617

🆔 @hackernewsgithubprojects

📦 smartcomputer-ai/lightspeed

Lightspeed: The Agent Harness That Actually Scales

Lightspeed is the deterministic agent harness that finally lets you run thousands of A.I. agents on a single machine without the cost blowing up. Most agent setups trap the brain inside a full operating system, meaning you have to spin up a whole virtual machine for every single task, even if it only needs three seconds. That is slow, expensive, and hard to secure. Lightspeed flips this by separating the agent's core logic from the actual compute. The thinking runs as a lightweight, durable workflow that survives restarts and stays cheap when idle.

📰 https://news.ycombinator.com/item?id=49817634

🆔 @hackernewsgithubprojects

📦 netease-youdao/confucius4-r2t2

The Speech Model That Never Backtracks

Confucius Four R-two-Two writes down words as they are spoken, then locks them in place, meaning the text never flickers or changes later. Most live captioning systems constantly rewrite what they just said, which feels janky and makes it impossible to feed that text into other tools instantly. This model from NetEase solves that by committing to its output, giving you a clean, stable stream of text with almost no delay. It’s built for people who need real-time subtitles or live translation that actually works without constant corrections. It’s fast, accurate, and open, making it a surprisingly solid choice for anyone building real-time audio apps.

🆔 @hackernewsgithubprojects

📦 prismorsec/prismor

Prismor: The Security Guard for AI Agents

Prismor intercepts AI agent tool calls in real time to block dangerous actions like secret leaks or prompt injections before they execute. It acts as a self-hosted control plane that sits between your coding agent and your machine, inspecting every command to catch malicious redirects or destructive mistakes. The project is agent-agnostic, meaning it works with Claude, OpenAI, and LangChain without needing you to rewrite your entire workflow. It runs a local dashboard where you can switch between just watching the agent or actively blocking it, giving you the power to audit what your AI is actually doing.

🆔 @hackernewsgithubprojects

📦 netease-youdao/confucius4-t3po

Confucius4-T3PO: Real-Time Translation Without Waiting

Confucius4-T3PO translates text the instant it appears, rather than waiting for a full sentence to finish. It works like a live interpreter, deciding for every tiny chunk of input whether to pause for more context or immediately output a translation. This model uses a clever reinforcement learning trick to balance speed and accuracy, so you get high-quality results without the lag of traditional tools. It is built for real-time apps where waiting is not an option, and it even keeps its general smartness intact for future upgrades. If you are building live translation features, this is the engine that makes it feel instant and natural.

🆔 @hackernewsgithubprojects

📦 colingpt9/clips-studio

Clips Kitty: AI Clipping on Your PC

Clips Kitty is the open-source clip generator that keeps your footage on your own computer. Instead of uploading long streams to a cloud server to find highlights, this app runs a local artificial intelligence model directly on your machine. It scans your videos to find the best moments, automatically crops them to a vertical format, and even tracks who is speaking to keep them centered in the frame. This means you can turn hours of content into ready-to-post shorts without paying any subscription fees or uploading sensitive data. It is built specifically for creators who want full control and no hidden costs.

📰 https://news.ycombinator.com/item?id=49504646

🆔 @hackernewsgithubprojects