📦 patchy631/jev-as-judge

Jev-as-a-Judge: Testing AI Refund Replies

Test whether an AI support bot is lying about refunds using a single, auditable evaluation run. This project, called Jev-as-a-Judge, solves the problem of trusting automated customer service by replaying a frozen set of tricky support chats against a language model judge. It is interesting because it prioritizes transparency, recording every raw answer and model version so you can see exactly how the verdict was reached. Instead of just giving a pass or fail, it breaks the reply down into specific checks for factual grounding and honesty. This makes it a practical, offline-friendly tool for anyone wanting to verify AI behavior without building a complex infrastructure.

🆔 @hackernewsgithubprojects

📦 yfyeung/prunedctc

Pruned CTC: Smarter Speech Training

Pruned CTC is the training tool that finally stops your computer from running out of memory when teaching speech models to recognize thousands of different words. The real surprise here is that it calculates the exact same correct answer as standard methods but uses far less space. Instead of storing a massive table of every possible word for every moment in time, it only looks at the specific words you are trying to learn right now. It does this by breaking the calculation into tiny pieces and throwing away the heavy data as soon as it is done.

🆔 @hackernewsgithubprojects

📦 vincemakes/kiso

kiso: The AI Agent That Remembers Everything

The most surprising thing about this AI agent is that it actually survives getting killed. Most coding assistants you run in your terminal vanish the instant your laptop crashes or you close the window, losing all your work in the process. This project solves that by writing every single step of the conversation to disk as it happens. So if your computer dies mid-task, you can just tell it to resume, and it picks up exactly where it left off without repeating any steps that already worked. It is essentially a durable memory for your AI helper, ensuring that a power outage doesn't wipe out hours of code.

🆔 @hackernewsgithubprojects

📦 jaydendavisnc/inkwave

Inkwave: The Browser Game That Builds Itself

Paint your team's color across a wet plaza while swimming through your own ink to refill and dodge. That is the core of Inkwave, a four versus four turf war shooter that runs entirely in your web browser without needing any heavy install. The most surprising part is that every single piece of the game is generated by code. The characters, the sound effects, the music, and even the textures are all procedurally created in real time. There are no downloaded asset files to load, which makes the whole experience remarkably light for a full 3D shooter.

🆔 @hackernewsgithubprojects

📦 mem0ai/openmemory

OpenMemory: Keep your AI coding context

OpenMemory is the open source tool that finally stops your AI coding work from getting stuck when you switch tools. It lets you move your actual chat history between different AI coding assistants like Claude Code and Codex in one go. Instead of losing your progress, you simply pick up right where you left off with all your context intact. It works by reading your old sessions and writing them into the new tool so you never have to start from scratch again. You can choose exactly which conversations to move, keeping your workflow smooth without any manual copying.

🆔 @hackernewsgithubprojects

📦 tarquinen/opencode-vim

opencode-vim: Vim for OpenCode

Your OpenCode terminal just gained the editing power of Vim, meaning you can now navigate your entire chat history and edit prompts without ever touching your mouse. This plugin brings familiar keyboard shortcuts directly into OpenCode, letting you scroll through messages, copy text, and refine your inputs using standard commands. It solves the friction of switching between your cursor and keys by keeping your hands on the home row, which is a huge win for speed and muscle memory. If you use OpenCode daily, this small update makes your workflow significantly faster and more precise.

🆔 @hackernewsgithubprojects

📦 makevoid/motion-graphics-music-video-skill

motion-graphics-music-video-skill

Turn your audio tracks into high-quality, fully edited motion graphics videos with just a single text prompt. This clever Claude Code plugin coordinates an entire production crew directly from your terminal. It acts as an automated director that analyzes your song's beats, maps out character references to keep faces consistent across scenes, and coordinates video generation models to animate everything perfectly. Local scripts handle the heavy lifting, using Python for audio mixing and Swift on macOS for blazing-fast video effects. If you make music and want professional videos without opening a complex editing timeline, install this plugin and let the agent build a masterpiece for you.

🆔 @hackernewsgithubprojects

📦 ryleighnewman/inputconfig

InputConfig: Turn Any Controller into a Mouse

InputConfig turns any game controller into a full keyboard and mouse on your Mac, letting you play unsupported games or drive your desktop without touching a single key. You can map buttons to type whole words, move the cursor, or even trigger Siri Shortcuts, and it works with everything from Xbox pads to MIDI keyboards. The coolest part is accessibility: you can run your whole computer from just one stick, making it a lifeline for people with limited mobility. No drivers, no subscriptions, and no data collection, it just works.

🆔 @hackernewsgithubprojects

📦 jwuthri/selfjev

Selfjev: The AI That Just Says Yes Or No

A tiny language model can answer complex questions without writing a single word. Selfjev is an open decisions model that skips the usual chatbot style entirely. Instead of generating text, it runs a quick check through a four billion parameter brain on a single graphics card to give you a typed answer with a confidence score. You ask if something is true, which option is best, or rate a score, and it just outputs the result. No essays, no fluff, just the decision. It solves the problem of needing a fast, deterministic answer from an AI without the slow, expensive process of generating every token.

🆔 @hackernewsgithubprojects

📦 levy-street/pokemon-claude-red

Pokemon Claude Red: The Game Made Entirely of Code

Pokemon Claude Red is the browser game that proves you don’t need image files to make a pixel art masterpiece. It is a full remake of the original Pokemon Red, built by an AI that draws every sprite, tile, and musical note directly in code. There are zero image files here; the game calculates the visual look of everything on the fly. You get the complete story, all one hundred and fifty-one Pokemon, and the classic glitches, all running right in your web browser. It is a wild look at what artificial intelligence can do when you ask it to build a world from scratch.

🆔 @hackernewsgithubprojects

📦 peterfriese/system-one-foundation-models

System One for Apple Foundation Models

Plug structured logic directly into Apple's native AI framework to get instant, guaranteed answers. This library connects a decision model to the Foundation Models API, so instead of asking a chatbot to guess, you define strict rules and get a definitive result in under two hundred milliseconds. The standout feature is the ability to run this logic entirely on your phone without touching the internet, using the device's built-in neural engine to keep data private and latency near zero. It transforms fuzzy AI text generation into precise, typed code that developers can trust for critical tasks like email triage.

🆔 @hackernewsgithubprojects

📦 peterfriese/jev-foundation-models

System One for Apple Foundation Models

Plug in a decision engine to Apple's native AI framework without the guesswork. This bridge connects System One models to your Swift code so you get a definitive answer, not a hallucinated guess. It evaluates your data in milliseconds and returns a typed result with a confidence score, letting your app automatically route tasks or escalate to humans when the model isn't sure. Whether you run the model entirely offline on your phone or hook it to the cloud, it fits right into your existing code. No more parsing messy text outputs, just clean, reliable decisions.

🆔 @hackernewsgithubprojects

📦 alphanu1/daytona-arcade-recomp

Daytona USA: The Arcade Game Rebuilt From Scratch

Daytona USA is the arcade racer that ditched emulation entirely to run the original hardware code directly on your computer. Instead of simulating old chips, developers took the game's processor instructions and translated them into modern, portable code, while the graphics engine builds every frame from scratch using your own video card. This means the car, the track, and the sound board all run natively on your desktop, delivering a speed and accuracy that traditional emulators simply cannot match. It is a rare look under the hood of a classic machine, proving that a nineteen-nineties racing game can still feel brand new when rebuilt from the ground up.

🆔 @hackernewsgithubprojects

📦 wenyiwu0111/rsigame

RSIGame: The Game That Builds Itself

A computer program can now build a playable video game, find its own bugs, fix them, and keep getting better on its own until it hits a limit. This is how RSIGame works. It uses AI to play the game, watch for glitches, and repair the code in a loop. The coolest part? It doesn't just guess what’s wrong. It records evidence, like screenshots and logs, to pinpoint exactly where the game breaks. Then, it fixes that specific issue, checks the whole game again to make sure nothing else broke, and saves the best version.

🆔 @hackernewsgithubprojects

📦 quimmedes/cafe-llama.cpp

Run Massive AI Models on Your Laptop

Run massive AI models on your laptop without expensive hardware. Cafe Llama C P P is an experimental engine that takes standard open source models and lets them breathe life into your computer. It solves the problem of memory limits by intelligently moving heavy parts of the model to your regular system memory or even your hard drive when your graphics card runs out of room. This means you can run complex reasoning tasks locally with surprising speed. It is a powerful way to keep your data private while still accessing advanced intelligence. Give it a try if you want to build AI apps that run anywhere.

🆔 @hackernewsgithubprojects

📦 chinesezjc/cordisclaw

CordisClaw: Plugins That Prove Themselves

CordisClaw builds a Rust runtime that treats plugins like strict legal contracts rather than loose code snippets. The most surprising thing it does is verify every plugin against machine-readable documentation before it ever runs. This means the system knows exactly what each piece of code can do and who it is allowed to talk to, preventing silent failures and hidden dependencies. It is a deliberate move away from messy dynamic loading toward a fully auditable, fail-fast architecture. If you are frustrated by plugin systems that break for unclear reasons, this project shows how making rules explicit turns chaos into clear, debugable logic.

🆔 @hackernewsgithubprojects

📦 romancypherpunk/12axes

12 Axes: A Political Quiz That Refuses to Simplify

Place yourself on a political map that refuses to flatten your complex beliefs into a single left or right label. The project called Twelve Axes measures twelve independent dimensions separately, from economic control to social morality, so you can see exactly where your views clash or align without averaging them out. By comparing your specific combination against hundreds of ideologies, historical regimes, and political figures, it reveals the unique tensions in your worldview that standard tests hide. It is a fascinating way to understand exactly where you stand relative to the world, offering a detailed portrait of your political identity in plain English.

🆔 @hackernewsgithubprojects

📦 continuum-ai-corp/orcapromptvault

OrcaPromptVault: The Hidden Rules of AI

Peek behind the curtain of your favorite AI tools. OrcaPromptVault preserves the exact instructions that power chatbots like ChatGPT, Claude, and Grok. It solves the mystery of why an AI behaves a certain way by archiving the specific text sent before you type a single word. Instead of guessing, you can verify exactly what the machine is told to do. This turns a black box into a clear, readable document. No more folklore about hidden rules, just the facts. Check it out to see how these systems really think.

🆔 @hackernewsgithubprojects

📦 chenxiachan/paperfold

PaperFold: Zoom Into Any Research Paper

PaperFold is the local paper reader that lets you zoom from full text to a single-word topic map. It solves the headache of holding a complex argument in your head by folding sections into concise summaries while keeping your place. As you pinch to zoom, words slide smoothly to their new spots so you always know exactly where you are. It even highlights your notes as you zoom out. If you read academic papers, this tool makes the structure instantly clear without losing the original detail.

🆔 @hackernewsgithubprojects

📦 coral79/pami-code

PAMI: Text to Human-Object Interaction

Turn text descriptions into realistic full-body human-object interactions. PAMI generates coordinated motion where people pick up, lift, or move objects without drifting or clipping through them. The key innovation is part-anchored motion, which tracks object position relative to specific body parts rather than treating them as separate entities. This approach solves common problems like missed contacts and penetration that plague previous methods. By using a coarse-to-fine generation hierarchy, PAMI first creates a general interaction from text, then refines it with detailed surface sensing for accurate contact geometry. The result is more faithful and physically plausible human-object interactions than existing approaches.

🆔 @hackernewsgithubprojects