📦 superxslam/tartanimu

TartanIMU: One Model for All Robots

TartanIMU is the single model that tracks any robot or human using only an inertial sensor. Usually, tracking algorithms need to be built separately for each type of movement, like a car or a dog. This project changes that by training one network on over one hundred hours of data from cars, quadrupeds, drones, and humans. It learns to predict three-dimensional velocity just from acceleration and rotation data. This means one set of weights can handle all these different bodies without needing specific settings for each. If you work on robotics state estimation, this is the universal positioning tool you have been waiting for.

🆔 @hackernewsgithubprojects

📦 alchaincyf/loop-engineering-orange-book

Loop Engineering: Stop Prompting, Start Building

Stop typing every single instruction into your AI assistant and start designing a system that runs itself. That is the core idea behind Loop Engineering, a new free guide from HuaShu that explains how to move beyond manual prompting. Instead of you being the human in the loop, you build a structure that wakes up on a timer, spawns its own helper agents, checks their work, and remembers what happened last. You build it once, and it keeps working without you hovering over the keyboard.

🆔 @hackernewsgithubprojects

📦 xflare-bot/pokemonlive

Pocket Battle Lab: AI-Generated Pokemon Battles

Pocket Battle Lab is the browser game that generates your Pokemon animations on the fly. It runs a standard turn-based battle where the rules engine handles all the damage and status effects, while artificial intelligence creates fresh, anime-style video clips for every single attack. Instead of looping the same old pre-recorded files, the system plans a short cinematic storyboard and renders it right there in the battlefield while the previous move is still playing. This local prototype proves that you can build a fully playable battle without ever opening a video editor. It is a fascinating look at how far we can push real-time content generation for retro-style games.

🆔 @hackernewsgithubprojects

📦 seunggeunkimkr/genuine-any-order

How AI Writes Code Any Way It Wants

Genuine Any Order is the code generation model that refuses to write programs strictly from left to right. It is built to prove that artificial intelligence can think about code structure just like a human would, deciding where to start and how to fill in the blanks. The system actually grows its own sequence of words as it works, rather than sticking to a fixed length, which lets it tackle complex Python scripts with surprising flexibility. This matters because it shows that machines can reason about the shape of a solution before they even finish writing it.

🆔 @hackernewsgithubprojects

📦 sf-stav/velogb10

veloGB10

NVIDIA’s DGX Spark has a hard ceiling of about two hundred and fifty-five gigabytes per second, and veloGB10 squeezes every drop out of it. This is a Rust inference engine built specifically for that hardware, letting you run massive language models at speeds that rival much bigger clusters. The coolest part? It uses multi-token prediction to guess several words at once, verifying them in a single flash. Because the math is identical to separate guesses, the result is perfectly lossless, not just an approximation. It even lets two machines work together to make one request faster, not just to fit the model.

🆔 @hackernewsgithubprojects

📦 alibaizhanov/mengram

Mengram: AI Memory That Learns From Mistakes

Mengram gives AI agents a memory that actually learns from its own mistakes, not just a log of what happened. It splits memory into three types: facts, events, and workflows. The coolest part is the procedural memory. When an agent fails at a task, like a deployment crashing because it forgot a step, Mengram doesn't just record the failure. It automatically updates the workflow to include the missing step so the agent never makes that mistake again. It’s like giving your AI a notebook where it writes down its own lessons learned. This means your agents get smarter over time without you having to manually tweak their instructions.

🆔 @hackernewsgithubprojects

📦 virtualluoucas/structab

StrucTab: The Smart Way to Read Tables

StrucTab is the table-parsing framework that finally understands the layout of complex spreadsheets. It stops treating a table image as a simple picture and instead breaks the job into three distinct steps: counting rows and columns, mapping out merged cells, and then generating the actual code. This approach fixes a common problem where models guess the structure and get it wrong, leading to messy, broken data. By using a smart scoring system that checks these steps separately, it ensures the output is accurate and structured properly.

🆔 @hackernewsgithubprojects

📦 disler/fixing-smartass-opus-5

Fixing Smartass Opus 5

Turn your smartest AI coding assistant into a precise engineering partner with a single markdown file. This project stops the model from burying simple answers under layers of fluff, fake enthusiasm, and unnecessary formatting. Instead of repeating the same verbose tics every time, the system prompt enforces a strict operating contract that demands clarity and cuts the noise. You get direct, actionable answers that respect your time and your output limits. It is a practical fix for anyone tired of scanning through pages of text just to find one useful sentence. Stop paying for tokens you will never read and start building with an agent that actually works.

🆔 @hackernewsgithubprojects

📦 breko861-hash/sol-luna-codex-orchestrator

Sol Luna Codex Orchestrator

Orchestrate AI coding agents without wasting money. Sol Luna Codex Orchestrator lets a main AI plan while cheaper workers handle specific tasks. It stops the expensive brain from doing grunt work. Instead, it sends small jobs to quick, low-cost models. This cuts costs and speeds up development. You get smart decisions without paying premium rates for everything. It is a simple template for efficient coding. Stop overpaying for AI. Use the right brain for the job. Try it and save cash.

🆔 @hackernewsgithubprojects

📦 mardlib/ensoniq-eps-16-plus

Ensoniq EPS-16 Plus: The 1990s Sampler, Reborn

The Ensoniq EPS-16 Plus emulator runs the original 1990s operating system on modern Mac hardware. Instead of simulating the sound with a modern sample player, this project builds a digital replica of the physical circuitry. It presents the authentic software to a virtual machine that mirrors the original chips, buttons, and floppy drives. This means the instrument behaves exactly like the vintage hardware you remember, complete with its specific quirks and workflows. You get the full, unaltered experience of the classic sampler without needing the aging original equipment. It is a precise technical recreation that preserves the unique character of this iconic machine for today’s musicians.

🆔 @hackernewsgithubprojects

📦 vavo/lora-pilot

One Container For All AI Training

A single Docker container can now run three separate Stable Diffusion training systems side by side without crashing. That is the core idea behind Lora Pilot, a prebuilt environment that solves the headache of managing conflicting software versions for custom AI model training. Instead of installing each tool separately and hoping they work together, this project bundles the trainers, visualizers, and image generation tools into one shared workspace. Everything saves to the same folder, so your work survives restarts and you never have to reconfigure paths.

🆔 @hackernewsgithubprojects

📦 lostb1t/remux

Remux: The Media Server That Doesn't Need Downloads

Remux streams movies directly from torrents without ever downloading the file to your hard drive. It is a self-hosted media server written in Rust that speaks the same language as Jellyfin, meaning your existing apps work with it immediately. The project combines local files, remote sources, and torrent streaming into a single library, so you can browse and play content without juggling multiple services. Because it handles the streaming process internally, you get instant playback with full support for subtitles and audio tracks. This approach keeps your storage clean while still giving you access to a vast array of media.

🆔 @hackernewsgithubprojects

📦 cswry/vosr

VOSR: The AI That Fixes Blurry Photos Without Text Prompts

VOSR is the image upscaler that restores sharpness without needing a text description. Most AI tools that fix blurry photos require you to type a prompt explaining what is in the picture, but this model relies entirely on its own visual understanding. It was accepted into the top computer vision conference for 2026, proving that you do not need language instructions to get high-quality results. By using a specialized neural network, it takes a low-resolution image and rebuilds the missing details directly from the visual data. This makes it incredibly useful for anyone who wants to improve old photos or screen captures without worrying about writing the right words.

🆔 @hackernewsgithubprojects

📦 ylchen333/comfyui-vosr2

VOSR 2.0: One-Step Image Upscaling in ComfyUI

ComfyUI-VOSR2 is the ComfyUI node that makes high-quality image upscaling feel instant. It wraps the VOSR 2.0 model, which uses a single-step diffusion process to turn low-resolution images into sharp, detailed versions without the usual multi-pass waiting. The coolest part is how it handles text and fine structures, making blurry letters and intricate patterns readable in one go. You just drop your image in, and it manages the heavy lifting, downloading the necessary model files automatically on first use. It’s perfect for anyone who wants professional-grade upscaling without complex setups.

🆔 @hackernewsgithubprojects

📦 grpcer/ownmem

OwnMem: Git-Native Memory for AI Coding Agents

An AI coding agent that forgets your project context is a liability, so OwnMem treats memory like source code. It stores project knowledge in plain Markdown files that live directly in your Git repository, meaning your agent's memory is version-controlled, reviewable, and travels with your code. When you ask an agent to recall something, it performs a deterministic, local search that requires no internet connection or extra API calls, ensuring the same query always yields the same result. This approach solves the problem of opaque agent memory by making every fact inspectable and reversible.

🆔 @hackernewsgithubprojects

📦 staatsgeheim/mathkernel

MathKernel: The Truth Engine for AI Math

Let an AI math engine verify its own answers. Math Kernel is a Python library that gives large language models a way to do math they can actually trust. Instead of just guessing, it uses different calculation styles, like exact logic or standard numbers, to check its work. It tags every result with a trust label, showing exactly how much you can rely on it. This solves a big problem where AIs confidently get math wrong. It is useful for anyone building tools that need reliable numbers. If you want your AI to be honest about what it knows, this is how you do it.

📰 https://news.ycombinator.com/item?id=49592366

🆔 @hackernewsgithubprojects

📦 blessedrebus/krawl

Krawl: The Deception Server That Catches Hackers

Krawl is the web deception server that turns your infrastructure into a trap for attackers. It runs a fake version of your application that looks completely normal on the surface but is actually filled with deliberately weak security holes that anyone with a basic scanner can find. When a hacker probes these vulnerable spots, the system quietly logs their activity and can even ban their IP address across your entire network. The real win here is that you are no longer just blocking bad traffic; you are actively learning how attackers think by watching them walk into a trap.

📰 https://news.ycombinator.com/item?id=47571680

🆔 @hackernewsgithubprojects

📦 majoorwaldi/comfyui-majoor-omnicam

Majoor OmniCam: Direct Your AI Video Camera

Majoor OmniCam is the camera layout tool that lets you choreograph your shots before the computer ever renders a single frame. Instead of guessing how a video model will move its camera and hoping for the best, this plug-in gives you a simple three-dimensional stage where you can physically place a virtual lens and animate its movement. You can scrub through time, watch a low-fidelity preview of the motion, and then hand that exact path to any compatible AI model. It solves the headache of trying to describe complex camera moves in text prompts by turning them into visual keyframes.

🆔 @hackernewsgithubprojects

📦 kayforkind/reimagine-it

Reimagine It: Design From Your Own Data

A single command can transform a plain web page into a visually stunning, fully functional design system without ever inventing a single detail. This tool, called Reimagine It, solves a common frustration where generic design prompts turn every website into the same sleek, corporate template. Instead of guessing what looks good, it reads the actual text, numbers, and colors already on the page to build a layout that is unique to that specific content. It extracts real facts to create charts, typography, and art that are completely offline and ready to use immediately. This means the final design is not just pretty, but actually accurate to the source material.

📰 https://news.ycombinator.com/item?id=49504599

🆔 @hackernewsgithubprojects

📦 20000419/fauxnix

fauxnix: The Windows Bash Translator

fauxnix lets your AI agent write standard Linux commands that run perfectly on Windows without installing a virtual machine or WSL. Because large language models are trained heavily on Bash, they often write PowerShell code that looks correct but fails in practice. Fauxnix solves this by taking those familiar commands, like listing files or searching for text, and deterministically translating them into native PowerShell. The output mimics the familiar Linux format, giving you clean results and clear error messages instead of confusing Windows stack traces. This makes it a reliable bridge for agents working on your desktop.

📰 https://news.ycombinator.com/item?id=49592864

🆔 @hackernewsgithubprojects