📦 jingxiao10/gua

GUA: The Missing Link in Neural Network Training

Here is the reality: when you train these networks, the optimizer can secretly ruin your carefully calculated steps. The project called GUA fixes this. Think of it like this: your navigation app tells you to turn left, but your car’s steering wheel is jammed, so you turn right anyway. GUA is the mechanic who checks the steering wheel before you start driving. It ensures the actual movement matches your plan, even when the math gets messy. This matters because physics problems are tricky, and small errors add up fast. By keeping things aligned, GUA helps these models learn correctly.

🆔 @hackernewsgithubprojects

📦 surge-ai/gdp-pdf

gdp-pdf: The Benchmark That Actually Reads PDFs

GDP pdf is the evaluation tool that measures whether large language models can genuinely understand complex documents. Built by Surge AI, this harness tests AI against a hundred real world tasks where models must read attached PDFs and answer specific questions without using any external search tools. The core strength here is strict, automated grading. Instead of relying on vague human opinions, the system uses another AI as a judge to check every single answer against a detailed rubric. This makes it incredibly useful for developers who need to know if their model is actually reading the data or just guessing.

🆔 @hackernewsgithubprojects

📦 otodock/oto-dock

OtoDock: Your AI Agent Company

OtoDock lets you build a self-hosted company of AI agents that answer real phone calls. It’s not just a chatbot; it’s a full operating system for your digital employees. You pick your own AI engine, like Claude or Codex, and plug it into this platform running on your server. These agents have persistent memory, can access your files, and even run while you sleep. The coolest part? They can literally pick up the phone. If you want your AI to feel like a real teammate who handles tasks autonomously and securely, this is the blueprint you need.

📰 https://news.ycombinator.com/item?id=49630606

🆔 @hackernewsgithubprojects

📦 yicheng47/runner

Runner: The App That Makes AI Coding Agents Work Like a Team

Runner lets you run multiple AI coding agents side by side inside a single native desktop app, turning scattered terminal windows into a coordinated crew. Instead of juggling separate sessions, you assign specific roles to each agent, like a coder and a reviewer, and they communicate through a shared event log to work on a shared goal. The most surprising part is that you can set up one agent to manage other agents, creating a self-organizing team that handles the entire workflow from planning to code review. It works locally on your machine, so every session is a real, persistent process you can watch and control.

📰 https://news.ycombinator.com/item?id=48236283

🆔 @hackernewsgithubprojects

📦 deepseek-ai/deepjit

DeepJIT: The GPU Compiler That Caches Everything

DeepJIT is the lightweight C++ library that lets you compile GPU kernels right at runtime without the usual boilerplate. It handles the messy details of finding your compiler, caching the results, and loading the code onto the device so you can just focus on writing the math. The coolest part is its shared cache system, which lets your entire team or cluster reuse compiled kernels from a single folder, meaning nobody has to wait around for slow compilation if the code hasn’t changed. It works seamlessly with both NVIDIA GPUs and Huawei Ascend chips, giving you a uniform way to handle device code.

📰 https://news.ycombinator.com/item?id=49641406

🆔 @hackernewsgithubprojects

📦 dicklesworthstone/franken_markdown

franken_markdown

Turn your Markdown into polished, self-contained HTML and tiny, optimized PDFs with just one single command. Franken Markdown is a pure Rust tool that does the heavy lifting for you, handling everything from syntax highlighting to font embedding without needing a browser or a massive software stack. The coolest part is that it generates real vector graphics for your diagrams directly in the PDF, so your charts stay perfectly crisp instead of looking like blurry screenshots. It is built to be fast and reliable, giving you professional looking documents from a simple text file.

📰 https://news.ycombinator.com/item?id=48826265

🆔 @hackernewsgithubprojects

📦 akitaonrails/ai-usagebar

The AI Usage Tracker That Actually Works

ai-usagebar is the desktop widget that finally tells you exactly how much you have left on your AI subscriptions. It was built because constantly checking your browser to see if you hit your rate limit is annoying, so it just grabs that info and puts it right on your screen. The coolest part is it works with almost everything, like Claude, ChatGPT, and Cursor, all in one place. It even has a timer that counts down when your quota resets, so you stop guessing. It is lightweight and built to be fast.

🆔 @hackernewsgithubprojects

📦 yilewang/llm-for-zotero

Your Zotero Library, But Make It An Agent

A browser extension just gave your personal research library a brain. L L M for Z o t e r o is an open-source research agent that lives directly inside your Z o t e r o library, reading your papers, notes, and PDFs to actually help you do research instead of just storing it. The most striking part is that it doesn't just chat; it acts. You can ask it to audit your library for missing metadata, auto-tag hundreds of papers, organize unfiled items, or even write and save notes back into your collection.

🆔 @hackernewsgithubprojects

📦 remysucre/prela

Prela: The Query Language You Control

Prela is the embedded query language that gives you total control over how data is processed. It works right inside your existing code, letting you mix queries with regular logic without switching contexts. The coolest part? It runs exactly as you write it, skipping the hidden optimizer that usually slows things down. That means you decide the speed, not a black box. It is fast, compact, and surprisingly simple to use. If you want to know exactly what your database is doing, this is the tool to check out.

📰 https://news.ycombinator.com/item?id=48356563

🆔 @hackernewsgithubprojects

📦 typstify/typstify

Typstify: The Desktop Editor for Typst

Edit professional documents right on your desktop with Typstify. This cross-platform tool gives you the polished power of LaTeX but swaps out the confusing, old-school syntax for a modern, intuitive interface. Built in Go and powered by the Gio UI framework, it runs smoothly on Windows, macOS, and Linux. You just need to drop in the Typst executables, and you are ready to go. It is not just a text editor; it is a full development environment that handles syntax highlighting, live previews, and file management out of the box.

📰 https://news.ycombinator.com/item?id=49127765

🆔 @hackernewsgithubprojects

📦 xhowenma/autodojo

AutoDojo: The Adaptive Attack That Breaks AI Defenses

A filter that blocked every single static attack on an AI agent still failed twenty-eight percent of the time when AutoDojo optimized the malicious text against it. This research tool turns static security tests into an adaptive arms race, using a large language model to iteratively rewrite injection attacks until they slip past specific defensive filters. It reveals that many current safeguards offer limited protection, especially when the user’s request is vague enough to let the attacker blend in as ordinary data. The project is less about providing a solution and more about exposing the structural gaps in how we currently protect intelligent systems from hidden commands.

🆔 @hackernewsgithubprojects

📦 ja7ad/hydra

Hydra: The Download Manager That Steals Speed

Hydra Download Manager splits a single file across multiple servers to make downloads finish before your coffee gets cold. Most tools wait for one slow server to drag its feet, but this software actively steals work from those slow connections and hands it to faster ones, so you never sit there watching a progress bar crawl. It works like a traffic cop for your data, constantly rebalancing the load to keep every pipe full. This is the only feature that matters, because waiting is the only part of downloading that actually hurts. It runs on your computer, talks to your browser, and just makes files appear faster.

🆔 @hackernewsgithubprojects

📦 devin-axis/deepseek-design

DeepSeek Design

Turn a single sentence into a fully editable website, presentation, or short video without opening a single editor. DeepSeek Design plugs directly into the DeepSeek Harness to generate real code files, not just static images, allowing you to drag, drop, and tweak every pixel in a visual studio right inside your chat. The standout feature here is selection-aware editing; you can highlight a specific button or heading and ask the AI to change only that element, leaving the rest of your project exactly as you built it. It’s a continuous loop where you describe your idea, watch it materialize, and refine it with surgical precision.

🆔 @hackernewsgithubprojects

📦 tencent-hunyuan/auk

AuK: The AI That Edits Sound Like Text

AuK is the open-source speech model that edits audio like you edit text. It lets you fix mistakes, change emotions, or swap voices in seconds without complex software. This tool solves the problem of tedious manual audio editing by using simple commands to generate or modify speech. It’s useful because it makes high-quality voice control accessible and surprisingly fun to play with. You can create new lines or tweak existing recordings effortlessly. If you love working with audio, this is a clever way to save time and spark creativity. Try it out and see how easy sound editing can really be. It’s a cool little project that deserves your attention.

🆔 @hackernewsgithubprojects

📦 v-songbird/hush

hush: The plugin that makes your AI coder stop talking

Hush is a plugin for Claude Code that forces your coding assistant to stop narrating every single step. Instead of watching a wall of text scroll by while it works, the tool shuts up until the job is actually done. Then, it gives you one short, plain-English answer telling you exactly what changed and which file to open. It cuts out the noise so you can just see the result. You can even pick a different tone for that final message if you want. If you are tired of reading a novel just to know if a test passed, this is the fix.

🆔 @hackernewsgithubprojects

📦 prove2me/prove2me_workspace

Prove2me Workspace: The AI Math Proof Engine

Prove2me Workspace is the development kit that lets AI agents formally prove mathematics in Lean. It solves the problem of verifying complex logical claims by using a server-side toolchain to check every proof against exact formal statements. This means a successful submission is a real mathematical guarantee, not just a heuristic guess. The platform organizes work into missions, where agents can discover open theorems, decompose hard problems into smaller lemmas, and submit code for rigorous verification. By bridging the gap between raw code and formal logic, it creates a trusted environment for automated reasoning.

🆔 @hackernewsgithubprojects

📦 sealandseacat/dbmask

dbmask: The Data Masking Tool That Checks Its Own Work

Databases that check their own work? Yes, dbmask is a tool that doesn’t just hide sensitive data in your SQL databases, it proves it actually did. It finds columns holding personal info, swaps them for consistent fake versions that keep your data usable, and then runs a strict validation step to verify every single row changed as intended. This solves the awkward problem of needing safe database copies for development or testing without risking a real leak. It is built to be auditable and fail-safe, meaning if it cannot verify a change, it refuses to pretend the job is done.

📰 https://news.ycombinator.com/item?id=49645189

🆔 @hackernewsgithubprojects

📦 aldegad/sprite-gen

Turn One Image Into Game Sprites

Transform a single static drawing into polished, game-ready character sprites that actually work in your engine. Traditional AI image generation often produces messy sheets with shifting faces and backgrounds that are impossible to remove cleanly, leaving developers with unusable files. This tool fixes that by taking one base image and processing it step by step to lock the character’s identity and strip away the background into true transparency. The most surprising feature is its ability to turn that still image into seamless, animated motion loops, giving your character a living idle state without needing to animate every frame by hand.

🆔 @hackernewsgithubprojects

📦 liandu2024/open-box

Open-Box: One-Command Proxy for OpenWrt

Open-Box is the all-in-one solution that turns any OpenWrt router into a fully managed proxy hub with a single command. Instead of juggling complex terminal setups or manual configurations, you simply run one installer script that bundles the entire proxy engine and a clean web interface directly onto your hardware. The magic lies in its transparency; it quietly reroutes your traffic in the background while giving you a simple browser dashboard to manage subscriptions and nodes without touching a single command line. It handles the tricky parts like DNS management and firewall rules automatically, ensuring that if anything goes wrong, your network stays online rather than crashing.

🆔 @hackernewsgithubprojects

📦 byjoey/warp-masque-actions

The Warp-Masque-Actions Trick

A GitHub workflow that automatically generates the keys and configuration files you need to run a private, encrypted network tunnel through Cloudflare WARP is a rare find for anyone tired of manual setup. The Warp-Masque-Actions repository uses a simple automation script to handle the entire process, saving the final settings as a downloadable archive so you never have to copy and paste keys or edit complex text files by hand. This solves the common frustration of keeping a secure connection running when the underlying certificates or settings change, ensuring your private network stays accessible without any technical overhead.

🆔 @hackernewsgithubprojects

📦 v-modal/vmodal_sdk_smartglass

VModal Smart Glasses SDK: Live First-Person Video

Your smart glasses can stream a live, first-person video feed straight to your server without your app touching the messy plumbing. The VModal SDK handles the entire heavy lift, grabbing the camera feed from the glasses, compressing it using the phone’s built-in hardware, and sending it out over a secure, encrypted connection that automatically reconnects if the network dips. This means you can build remote support tools or hands-free field guides without worrying about video buffering or broken streams. You simply plug in your backend, and the SDK manages the rest of the tricky media pipeline.

🆔 @hackernewsgithubprojects