📦 omnijev/onejev

OneJev: The AI That Answers In One Shot

OneJev is the multimodal decision model that answers typed questions about screens, photos, and video in a single forward pass. You send it a screenshot or a video clip with a few specific questions, and it immediately returns a calibrated probability for every option without needing to write out a long chain of thought first. The whole screen is read once, and each question branches off that single look at the data. It’s incredibly fast, with the smaller versions answering ten questions about a screen in roughly the same time it takes to answer just one.

🆔 @hackernewsgithubprojects

📦 mavka-ai/unit-tests-skills

Unit Test Skills: Smarter AI Testing

Unit tests generated by AI agents now catch nearly double the hidden bugs compared to standard prompts, even when coverage looks identical. This project, called Unit Test Skills, solves the problem where AI writes tests that pass but fail to catch real logic errors. Instead of just asking for code, it gives the agent a strict process to plan test cases first, then generate, compile, and run them. It works with tools like Claude Code and Cursor for Java projects. The result is tests that actually find defects instead of just checking boxes. If you use AI to write tests, this is the missing piece for real quality.

🆔 @hackernewsgithubprojects

📦 xiaomimimo/verl

verl: Train AI Models with Reinforcement Learning

verl lets you fine-tune large language models using reinforcement learning, which basically means teaching an AI by rewarding it for good answers. The cool part is its hybrid engine. Instead of stopping your training to generate new text, it swaps memory efficiently on the fly. This kills the bottleneck where computers wait around doing nothing. It is built by ByteDance for production use, handling massive models across hundreds of GPUs. If you are trying to make your AI sharper and faster, this is the tool to look at. It turns a complex training pipeline into something you can actually run without breaking a sweat.

🆔 @hackernewsgithubprojects

📦 ludattilo/revit-mcp-server

Revit MCP Server: AI Meets Architecture

Revit MCP Server is the AI bridge that lets you talk to your building models. It connects your favorite AI assistant directly to Autodesk Revit, giving it the power to read, create, and modify elements in real time. Imagine telling your AI to find all the doors, check for clashes, or even export a PDF, and watching it happen instantly. This tool turns your AI from a chatbot into a hands-on collaborator for your architectural work. It is not just about automation; it is about giving your AI a direct line to your design.

🆔 @hackernewsgithubprojects

📦 beamivalice/sushi

Sushi: The Local AI Engine Built for Apple Silicon

Sushi is a custom inference engine that lets a single Mac run massive artificial intelligence models locally without the cloud. Built specifically for Apple chips, it uses a clever compression method to fit huge models into your computer's memory while keeping the output surprisingly accurate. It acts as a drop-in replacement for standard AI servers, meaning your coding tools and chat apps can connect to it instantly. You get full privacy and blazing fast performance because everything runs right on your hardware. No waiting, no data leaks, just pure speed. If you want to own your AI experience, this is the engine to watch.

🆔 @hackernewsgithubprojects

📦 twosevenonet/injectsetconsole

InjectSetConsole

Windows process injection can now be achieved completely without using the typical memory writing functions that security software constantly monitors. A proof of concept called injectsetconsole demonstrates how to sneak code into a target program by taking advantage of standard interactive console inputs. Instead of forcing data directly into a remote application's memory space, it redirects standard input to a named pipe and simply writes the payload as if typing a regular command. The target process itself reads this data and stores it in its own memory. Security tools looking for typical injection signatures get bypassed because the target application does the heavy lifting.

🆔 @hackernewsgithubprojects

📦 beatapi/awesome-jev

Awesome JEV: The Curated Guide to Agent Logic

JEV is quietly becoming the brain behind some of the smartest coding agents, and Awesome JEV is the map that shows you where. It is not a code library or a heavy framework, but a rigorously reviewed catalog of real projects that actually use this technology. The most interesting part is that every single entry includes direct links to the specific lines of code where the decision logic happens, so you are not just reading a description; you are seeing the actual implementation. This solves the massive problem of finding trustworthy examples in a sea of marketing fluff and vague demos.

🆔 @hackernewsgithubprojects

📦 deepseek-ai/deepgemm-ascend

DeepGEMM Ascend: NPU Matrix Magic

DeepGEMM Ascend takes the painful, low-level mess of writing high-speed math kernels for Huawei Ascend chips and hides it behind a simple, clean interface. Normally, squeezing maximum speed out of this hardware means wrestling with confusing memory layouts and complex alignment rules, but this library wraps those headaches in a neat package. It lets you run heavy AI math, like matrix multiplications for large language models, at speeds that hit the absolute physical limit of the chip. Because it uses the same commands as the original DeepGEMM project, you can swap it in without rewriting your code.

📰 https://news.ycombinator.com/item?id=49903881

🆔 @hackernewsgithubprojects

📦 vivy-yi/awesome-agent-orchestration

Awesome Agent Orchestration: The Ultimate AI Agent List

Coordinate a whole team of A.I. workers to tackle complex jobs you could never handle alone. That is the power of the Awesome Agent Orchestration list, which maps out the entire landscape of frameworks like AutoGen and CrewAI. Instead of guessing which tool to trust, this curated guide organizes every major option from simple chat bots to massive swarm systems, along with the communication standards they all use. It turns a chaotic sea of options into a clear map, showing you exactly how these digital teams talk and work together.

🆔 @hackernewsgithubprojects

📦 antseed/antseed

Antseed: The P2P AI Network Without Gatekeepers

Antseed is the peer-to-peer network that turns AI inference into an open marketplace. Instead of routing every request through a single corporate server, this system lets users run their own nodes to sell AI services directly to others. You can plug in your own model or API key and share your spare computing power, while buyers automatically find the cheapest or fastest provider nearby. The result is a decentralized marketplace where you keep control over your data and your costs, removing the middleman entirely. If you want to own your AI infrastructure rather than renting it, this is the project to watch.

🆔 @hackernewsgithubprojects

📦 marspa73/atarijam

JAM: AI in 30 Kilobytes on an Atari 800

A working neural network language model now fits inside thirty kilobytes of memory on a 1979 Atari 800. This project, called JAM, proves that real generative artificial intelligence can run on ancient eight-bit hardware without needing a modern processor. It uses a deterministic neural network to generate text character by character directly on the machine, solving the problem of running complex AI on devices with almost no computational power. By squeezing a functioning brain into such a tiny footprint, it shows how far minimal computing has come. The result is a fascinating, albeit limited, demonstration of silicon intelligence running on a vintage gaming console.

🆔 @hackernewsgithubprojects

📦 hitheshkaranth/openterminalui

OpenTerminalUI: The Open Source Bloomberg Alternative

OpenTerminalUI is a self-hosted trading terminal that actually reads the company reports you are supposed to be reading, turning dense financial filings into verified, cited insights without hallucinating a single number. It works offline on your own hardware, giving you full control over your data and the AI models that crunch it. If you are tired of black-box algorithms that guess why a stock moved, this open-source tool lets you see exactly which quote in the annual report supports the claim. It is the difference between trusting a vibe and trusting a documented fact, all wrapped in a fast, keyboard-driven interface that respects your privacy and your setup.

🆔 @hackernewsgithubprojects

📦 echris6/motion-video-kit

The AI Video Quality Check

Motion Video Kit is the AI video toolkit that refuses to let you ship bad work. It packages a rigorous review loop where fresh critics grade every frame against strict quality standards, catching issues that automated tools miss. This prevents the common trap where AI video looks technically smooth but feels flat and generic. By forcing you to measure loudness, check for frozen frames, and verify motion physics, it ensures the final cut feels premium and intentional. You get a commercial that actually looks like it was made by a human studio. Stop guessing if your video is good. Let the critics tell you the truth before your clients do.

🆔 @hackernewsgithubprojects

📦 evilsheep/max-payne-2-vr

Max Payne 2 VR: Bullet Time in Headset

A two decade old noir thriller now features manual magazine reloading and physical grenade catching in virtual reality. Max Payne 2 VR is a free mod that lets you play as Max or Mona with tracked hands and room scale movement. It adds tactile combat where you physically swing weapons and catch live grenades mid-air. While it is a fan project for a classic PC game, it transforms the famous bullet time mechanic into a visceral, spatial experience. This is strictly for PC VR setups like Quest 3 or PSVR2. The takeaway is that even retro games can gain modern spatial depth through dedicated community mods.

🆔 @hackernewsgithubprojects

📦 iquestlab/modularrsi

Modular RSI: The Universal Test Harness for AI Agents

Modular RSI is a single command-line tool that lets you run the same AI coding agent on over sixty different real-world benchmarks. Instead of writing custom scripts for every new test suite, developers plug their agent into this universal harness. It handles the messy details like setting up secure sandboxes and checking if the code actually works. This means you can finally compare how well different AI models handle tasks like fixing bugs or writing new features, all using one consistent setup. It turns the chaotic process of testing AI agents into a simple, repeatable workflow.

🆔 @hackernewsgithubprojects

📦 private-garden-labs/garden-desk

Garden Desk: AI That Never Leaves Your Computer

Garden Desk runs powerful artificial intelligence entirely inside a locked virtual room on your own computer, meaning your files never touch the internet. It solves the privacy gap for people who need help organizing, reviewing, or analyzing their own documents without trusting a cloud server with their data. You simply pick a folder, describe what you need, and the software works inside that secure, offline bubble to produce results you can save wherever you like. This project proves that useful local automation is possible without sacrificing your control over your information. If you value your privacy, this is the tool to watch.

🆔 @hackernewsgithubprojects

📦 trevornk/ramblr

Ramblr: Dictate Anywhere on Android

Dictate text into any Android app with a single tap, completely offline if you choose. Ramblr is a voice input tool that transcribes your speech directly on your phone, so your data never leaves your device. The coolest part is the AI cleanup feature, which uses a local large language model to polish your messy dictation into clean, professional text before you even see it. It is not just a basic recorder; it actively fixes your grammar and tone right on your screen. This means you can talk naturally and let the software handle the editing, giving you total privacy and speed.

🆔 @hackernewsgithubprojects

📦 eifr/vigilo

Vigilo: Your Phone Is a Security Camera

Vigilo is the web app that turns your phone into a live security camera without any extra hardware. It works entirely in your browser, using computer vision to spot motion and track objects in real time, which means you do not need to install heavy software or connect to a private cloud. The coolest part is that you can set up alerts to send photos directly to your Telegram chat, so you get an instant snapshot every time something moves. It is a clever way to use the device already in your pocket as a simple, portable eye on your home or workspace.

🆔 @hackernewsgithubprojects

📦 opencoven/coven

Coven: The AI Runtime That Actually Remembers

Coven is the local-first runtime that finally gives your AI coding agents a real memory and a strict leash. Instead of letting a chatbot wander loose with your credentials, it keeps sessions locked to your specific project, ensuring the agent only does what you explicitly allow. The coolest part? It speaks every dialect. Whether you run Claude Code, Codex, or Copilot, Coven sits underneath them all, turning your scattered AI tools into one coherent, durable workflow that remembers your codebase context even after you close the terminal. It is less about flashy features and more about making sure your digital helpers actually stay in their lane.

🆔 @hackernewsgithubprojects

📦 menloresearch/isaac_asimov

Teach a Humanoid Robot to Walk

Train a real humanoid robot to walk by teaching it in a virtual playground. This project gives you the exact code to build walking brains for the Asimov one, a two-legged bot that looks like a tiny human. Instead of guessing how to balance, the software uses a clever trick to copy natural movement data, making the robot’s steps look smooth and lifelike rather than robotic. It handles the heavy math so you can focus on the fun part: watching your digital student stumble, recover, and eventually strut. Once it learns in the simulation, you deploy that same brain to the actual hardware.

🆔 @hackernewsgithubprojects