📦 sergiuszm/ninfer-4090

Run a 27B Model on One GPU

Ninfor forty ninety lets you run a massive twenty seven billion parameter language model on a single consumer graphics card, specifically an RTX 4090. Most people assume that running such a large model with its huge context window requires a rack of servers or a multi-thousand dollar workstation. This project solves that by squeezing the model into the available memory using a clever compression trick that stores the model's memory in a highly efficient format. It doesn't just run, it runs fast, hitting speeds that make real-time coding assistance actually feel responsive.

🆔 @hackernewsgithubprojects

📦 iwtbs27/collab-agent

Collab Agent: One Agent, Every Screen

Collab Agent makes a single A.I. brain work across your web browser, desktop apps, and command line at the same time. Instead of one bot stuck in one window, a team leader agent breaks your task into pieces and hands them to specialized workers who actually click, type, and code where they belong. The coolest part is that these workers pass real files and notes to each other, so if one gets stuck, another can pick it up. It is a lightweight system for getting complex digital work done without babysitting every step. If you want your A.I.

🆔 @hackernewsgithubprojects

📦 iiiiiiirene/learnibridge

LearniBridge: Faster AI Video

LearniBridge makes AI video generation significantly faster by fixing a critical flaw in how previous speed-up tricks handled data. Standard caching methods just copy old data, which causes blurry or broken images when you try to go really fast. This project learns a small, smart correction that fixes that damage using only three to five training examples. It skips heavy calculations entirely and replaces them with a quick, lightweight adjustment. The result is a massive speed boost for top video models without sacrificing quality. It is a clever way to make high-end AI tools run much faster on your hardware.

🆔 @hackernewsgithubprojects

📦 werwolv/riscv-emulator

riscv-emulator

Boot a full Linux kernel inside a C++ program you can actually read. This RISC-V emulator simulates a thirty-two-bit processor from scratch, handling memory and interrupts so real operating system code runs right in your browser or on your desktop. Instead of just running binaries, it lets you see exactly how the machine talks to the hardware, making it perfect for understanding how computers actually work. You get to witness the magic of virtual memory and page tables without needing a physical chip. It is the fastest way to learn how the low level bits and bytes fit together into a working system.

📰 https://news.ycombinator.com/item?id=49603397

🆔 @hackernewsgithubprojects

📦 mic-dkfz/fedsegnoisebench

FedSegNoiseBench: Testing AI on Messy Data

FedSegNoiseBench is the benchmark that finally tests medical AI against the messy reality of human error. In real hospitals, doctors rarely agree on exactly where a tumor starts and ends, creating noise that breaks standard training. This tool fixes that by pairing six diverse datasets with four distinct ways that data gets corrupted across different servers. Instead of relying on fake errors, it measures how well specific algorithms handle these real-world imperfections. The result is a clear guide showing which strategy actually works best when the labels are wrong. If you are building health AI, this is the map you need to find the path forward.

🆔 @hackernewsgithubprojects

📦 najmuzzaman-mohammad/gawkbot

gawkbot: The AI That Builds Its Own Dashboard

Turn your manual chores into live, interactive dashboards with gawkbot. It grabs a workflow you describe and builds a dedicated mini-app to run it automatically on your own computer. The coolest part is the approval gate: it handles the reading and thinking, but it waits for your click before touching any external system. This gives you a false sense of control, which is exactly the point. Your bots work around the clock, but you stay in the loop. It’s a clever, self-hosted way to automate the boring stuff while keeping your hands on the wheel. Check it out.

🆔 @hackernewsgithubprojects

📦 maxhu-xuan/artifactproof

Artifactproof: Verify your PowerPoint integrity offline

A tiny Python tool lets you cryptographically prove that the PowerPoint file you are about to send is the exact same one that passed quality control. Artifactproof creates a signed receipt bound to the specific file contents and structural checks you ran previously. If anyone swaps the file, changes a slide, or tampers with the evidence, the tool instantly detects the mismatch. Everything runs locally on your machine without an internet connection, ensuring your data never leaves your hands. It is a simple, offline guard against accidental or malicious file swaps right before delivery.

🆔 @hackernewsgithubprojects

📦 ljx-chase/research-field-onboarding

Field Onboarding: The AI Skill That Teaches You Step By Step

Field Onboarding is an AI skill that refuses to lecture you until it knows exactly what you already understand. Most assistants blast you with twelve unexplained terms, leaving you stranded, but this project forces the computer to pause and ask where you stand first. It builds a personalized ladder of concepts, teaching only one rung at a time and checking your understanding with real diagnostic questions rather than vague prompts. It also guarantees that every paper it cites is either verified in the moment or explicitly labeled as unverified, so you never waste time chasing a ghost reference.

🆔 @hackernewsgithubprojects

📦 benyronald77/uajy-academic-rag-chatbot

The campus chatbot that never lies

Stop guessing which university rule applies to your scholarship. This project builds a chatbot that reads the official academic handbook for Universitas Atma Jaya Yogyakarta so you get straight answers. It scans the full PDF, breaks it into pieces, and stores them in a local index. When you ask a question, it finds the exact matching text and sends only that to the AI. The result is an answer that cites the specific page number, so you can verify it yourself. If the document doesn't contain the answer, the bot admits it instead of making something up.

🆔 @hackernewsgithubprojects

📦 yean-sec/strikeagent_atkbrain-flash

StrikeAgent: The AI Pentester That Learns From Every Attack

StrikeAgent is the AI penetration testing platform that actually remembers how to break things. Developed by the Yean team, it stops treating hacking like a simple script and instead builds a living attack map. The standout feature is its self-evolving loop. After every scan, the system distills successful techniques into a shared memory library. This means the next attack starts smarter than the last one. It is not just guessing; it is learning. The team focused heavily on accuracy, adding a double-verification step to ensure reported vulnerabilities are real. It turns chaotic security testing into a precise, improving process.

🆔 @hackernewsgithubprojects

📦 openwam-official/openwam

OpenWAM: The Modular Robot Brain

OpenWAM has trained a single robot policy on six thousand hours of human and robot video. It is a free toolkit for building world action models, which are AIs that learn to move by understanding how the physical world behaves. Instead of treating the AI as a black box, it breaks the system into modular parts so you can swap out different video backbones and test them against eight major robot benchmarks. This open approach lets researchers see exactly how the model learns to connect what it sees with the actions it takes. It is a serious, systematic way to build robots that actually understand their environment.

🆔 @hackernewsgithubprojects

📦 kjgx66f/usque-custom-pro

Usque MASQUE Pro: One-Click WARP Config Tool

Usque MASQUE Pro is the browser tool that turns complex Cloudflare WARP setup into a single click. It solves the headache of manually editing technical files for secure networking by registering your device and generating compatible settings for major clients like Clash and Shadowrocket directly in the browser. The standout feature is its smart traffic routing, which automatically directs specific apps like Netflix or GitHub to the best available node while keeping other traffic direct. This removes the guesswork from configuration and ensures your connection works without constant tweaking. If you want a clean, automatic setup for your proxy tools, this is the practical choice.

🆔 @hackernewsgithubprojects

📦 buzzkit-dev/buzzkit

BuzzKit

BuzzKit is the open source notification orchestrator that handles the entire journey of a message for you. It solves the messy problem of connecting directly to every single delivery service by acting as a central hub. You just report user events, and the built in workflows decide exactly what to send and when. This means you get automatic retries, scheduling, and delivery tracking without writing custom code for each channel. It currently supports iOS push notifications, with email and Android coming later. For developers who want reliable messaging without the boilerplate, this is a powerful, self hosted option worth watching.

🆔 @hackernewsgithubprojects

📦 forsy-ai/biosecurity-agent

Biosecurity Agent: Your AI Guardian

Track the biological health of anything you care about. This local tool turns any target, like your family or a specific plant, into a live security world. It watches news, scientific papers, and public data to spot threats before they happen. The most surprising part is how it simulates the future. It predicts where a risk might travel in the coming days, then recommends specific actions to stop it. Everything runs on your machine, keeping your data private while giving you a clear, evidence-based view of what could go wrong. It is not just about monitoring; it is about preparing.

🆔 @hackernewsgithubprojects

📦 huawei-bayerlab/marigold-v2

Marigold V2: One Image, Full Depth

Turn any single photo into a full 3D depth map, and do it on your own laptop. Marigold V2 takes a standard image and reconstructs exactly how far every pixel is from the camera, capturing hair-thin details like fur and sharp edges. The coolest part? You can train the model in under a week on a single consumer graphics card, meaning no expensive server farm is required. It’s a practical, high-quality tool for anyone needing spatial data from flat pictures. Just point it at an image, and get a reliable 3D view. It’s surprisingly accessible for independent developers.

🆔 @hackernewsgithubprojects

📦 xzf-thu/voicemem

VoiceMem: The Voice Assistant That Actually Remembers

VoiceMem gives real-time voice agents a permanent memory that tracks not just what you say, but how you feel while you are saying it. The project splits this knowledge into two distinct streams, with one side holding factual details and the other capturing emotional context. The entire process happens in a continuous flow, meaning the system retrieves relevant memories before you even finish speaking. This architecture keeps response times incredibly low while using a tiny fraction of the data required by standard tools. It is a practical, fully open-source solution for building assistants that truly understand you over time.

🆔 @hackernewsgithubprojects

📦 empiricaai/empirica

Empirica: The AI Mirror

Empirica stops AI coding agents from confidently guessing by forcing them to prove what they know before touching your code. It acts as a measurement layer that tracks exactly how much the AI understands, blocking any edits until that understanding hits a specific threshold. This creates a live, visible score of confidence that updates in real time, ensuring the system only acts when it is actually ready. By remembering what it learned across different sessions, the tool prevents the frustrating repetition of dead ends. It turns invisible AI reasoning into a clear, measurable process you can watch.

🆔 @hackernewsgithubprojects

📦 roiamiel1/build-deep-learning-from-scratch

Build Deep Learning From Scratch: The PyTorch Reimplementation Course

Build Deep Learning From Scratch is the coding curriculum that makes you the autodiff library by forcing you to rebuild PyTorch internally. It bans every deep learning package and allows only raw array math and plotting, so you cannot import backpropagation. You start with a single number and manually trace how a tiny change ripples through the math to calculate the gradient. As you move up, you grow that simple calculation into full neural networks, image models, and language transformers. This hands-on approach ensures that every update inside a model is code you actually understand. It turns a mysterious framework into a machine you built yourself.

🆔 @hackernewsgithubprojects

📦 francozanardi/tscaps

Tscaps: Browser-Based Video Subtitles

Burn animated captions directly into your video right from your browser, completely free. Tscaps is an open-source editor designed specifically for short-form content like TikToks and Reels. It solves the privacy headache of cloud uploaders by running the entire transcription process locally on your device using Whisper. You get the polished, eye-catching text overlays without sending your footage to any server. The coolest part? Every caption style is built with pure CSS. That means if the built-in templates aren’t your vibe, you can write custom code to make the text dance exactly how you want. It’s the perfect privacy-first tool for creators who want total control over their visual style.

🆔 @hackernewsgithubprojects

📦 ultralytics/inference

Run YOLO at C Speed

Inference is the command line tool that brings lightning fast computer vision to your terminal without the Python mess. Written in Rust, it runs YOLO models directly on your hardware, skipping the heavy runtime setup that usually slows things down. You just point it at an image or video, and it detects objects, poses, and segments instantly, even downloading the model for you if you forget. It feels like cheating, letting you build real time vision apps that actually run on your laptop or server. If you want speed and simplicity, this is your new favorite tool.

🆔 @hackernewsgithubprojects