📦 antinomie-lab/pi-book

Inside the Pi Agent Loop

Pi Book is the architecture guide that actually shows you the code behind the agent loop. It is a workspace for a book that breaks down a specific library, explaining exactly how it handles tool execution and data streaming without becoming a bloated framework. The coolest part is that every single claim in the text is backed by a direct link to the exact line of source code, so you can verify every detail yourself without guessing. It is organized to build your mental model step by step, from the big picture down to the intricate parts.

🆔 @hackernewsgithubprojects

📦 inclusionai/areno

AReno: Train AI Models on Your Laptop

Train a large language model on your single local machine using AReno. This toolkit eliminates the need for massive server clusters or complex infrastructure, letting you run reinforcement learning, supervised fine-tuning, and even agentic workflows directly on your own hardware. It works by handling everything from data loading to the final training step in one self-contained package, so you are not stuck wiring together separate tools. You can even watch a small model learn to play a browser game from scratch, proving the system works end to end. The real value is accessibility. You no longer need a giant cluster to experiment with advanced AI techniques.

🆔 @hackernewsgithubprojects

📦 bytedtsinghua-sia/cuda-agent

This AI Beats Top Models at GPU Coding

Watch this AI agent outcode the best general purpose models at writing fast GPU code. Researchers trained a specialized agent to generate high performance CUDA kernels, beating advanced systems on difficult benchmarks. The project releases the training data and the exact workflow rules used to build it. It turns complex hardware optimization into a repeatable, automated process. This gives developers a proven blueprint for accelerating their own models. You get a clear path to faster code without years of manual trial and error. This is how the next generation of efficient AI training is being built. Check it out to see how agentic reinforcement learning is changing GPU programming.

🆔 @hackernewsgithubprojects

📦 perseus-computing-llc/mimir

Give Your AI Agent a Permanent Brain

Stop letting your AI agent forget everything the moment a session ends. Mimir gives your AI persistent, encrypted memory that lives locally on one file, completely offline. It uses military-grade encryption to keep your data safe, so your agent can remember past lessons and facts across different projects without sending anything to the cloud. It works with any tool, and because it is local-first, you keep total control. No servers, no subscriptions, just smart, private memory. If you want your AI to actually learn and stop repeating mistakes, this is the tool you need.

🆔 @hackernewsgithubprojects

📦 thanhhff/selectanytree

SelectAnyTree: Click to Segment Forests

One click on a 3D forest scan isolates a single tree, and it nails the accuracy target almost instantly. SelectAnyTree solves the nightmare of manually labeling millions of LiDAR points in dense woods, which is usually tedious and error-prone. It works like a smart paint tool where you simply tap a spot to select a specific tree, and the system instantly draws the outline. The clever part is that it processes the entire forest scene just once, so switching between different trees is fast because it doesn't have to re-analyze the raw data every time. It beats previous tools by a wide margin using far fewer computing resources.

🆔 @hackernewsgithubprojects

📦 michaelbf16/bedrock-rl

Train AI Agents in Minecraft with Bedrock RL

Bedrock RL is the modular framework that finally lets you train AI agents inside Minecraft with reproducible, verifiable results. It solves the headache of messy AI training by giving you a single command to generate synthetic data, check your setup, and run reinforcement learning experiments. The coolest part is its synthetic data generator, which creates realistic gameplay trajectories that you can replay in three different visual styles to prove your agent actually learned the task. You define your goal in simple config files, and the system handles the heavy lifting of data creation and model training.

🆔 @hackernewsgithubprojects

📦 mirros-lab/harnesseval-w

HarnessEval-W: Why AI Needs to Explain Its Grades

HarnessEval-W is the evaluation framework that finally makes AI world models show their work. Instead of just dumping out a single, opaque score for a generated video, this tool acts like a critical movie critic. It breaks down the footage into specific checks, asking if objects behave according to physics and if the environment stays consistent as the camera moves. The coolest part? It doesn't just guess. It builds a transparent evidence tree, showing exactly which frames and logical steps led to its final verdict. This means when a model gets a low score, you can actually see why it failed, rather than just shrugging at a number.

🆔 @hackernewsgithubprojects

📦 ayaan-lashari/nerftrack

NerfTrack: Your Local Codex Bill Checker

NerfTrack is the local desktop app that finally makes your AI coding assistant's hidden costs visible and easy to understand. It sits quietly on your computer, reading the usage records from your Codex account without ever sending your private code or credentials to a cloud server. The app turns those raw logs into a simple, weekly estimate of what you would have paid if you were using a standard API plan. This is a huge win for developers who want to budget their tech stack without guessing, offering a clear, private, and instant look at their actual spending.

🆔 @hackernewsgithubprojects

📦 ig-imanish/manixh

Manixh: The AI Portfolio Blueprint

manixh is the portfolio template that treats your AI coding agent like a professional architect. It ships with a set of strict markdown specs that tell the AI exactly how to build your site, ensuring it never invents weird designs or breaks your code. You run one simple command to install these rules, paste a prompt into tools like Claude or Cursor, and the AI builds a polished, dark-mode React portfolio for you. It is the perfect way to get a custom developer website without writing a single line of code yourself.

🆔 @hackernewsgithubprojects

📦 gaurav-gosain/tuios

TUIOS: The Terminal Window Manager That Actually Works

TUIOS is the terminal multiplexer that finally makes your command line feel like a real desktop. Built in Go, it lets you juggle multiple terminal panes with a vim-style interface, so you can manage code, servers, and scripts all in one place without tab overload. The coolest part is its smart tiling system, which automatically arranges your windows to fit your screen perfectly, just like a modern window manager. It even runs in the background, keeping your sessions alive if you disconnect. This is perfect for developers who want power and simplicity without leaving their terminal.

🆔 @hackernewsgithubprojects

📦 matinsenpai/senpaiscanner

SenPai Scanner Finds Working Cloudflare Nodes

SenPai Scanner digs through millions of Cloudflare addresses to find stable internet exits for users on restricted networks. It does not just check if a server is online; it validates if your specific proxy configuration actually works through it using a built-in Xray core. This two-step process ensures you only export servers that genuinely support your connection type, saving you from wasted time. The tool provides a simple interface to scan, rank, and copy working endpoints directly into your proxy client, making it a precise utility for maintaining reliable access in difficult network conditions.

🆔 @hackernewsgithubprojects

📦 jewbetcha/openflight

OpenFlight: DIY Golf Launch Monitor with Doppler Radar

OpenFlight is the do-it-yourself golf launch monitor that turns affordable radar hardware into a precision shot tracker. While commercial launch monitors are pricey and closed, this open-source project lets you build your own system using a Doppler radar module and a Raspberry Pi. The core magic happens when you hit the ball: a sound trigger detects the impact, firing the radar to capture high-speed data. From that, the software calculates your ball speed, club speed, and smash factor. It even estimates carry distance using a ballistic model, so you get useful feedback even without the optional angle radar.

🆔 @hackernewsgithubprojects

📦 llm-as-a-verifier/turboagent

Turbo Agent: The AI Proxy That Judges Its Own Answers

Turbo Agent is the local proxy that makes your AI coding tools sharper by having the machine judge its own work. Instead of blindly accepting the first answer a large language model gives, this tool sits between your application and the server, sending multiple requests at once. It then uses a tournament-style method to compare the results and picks the most reliable one, effectively filtering out the noise so you get a better answer without changing your setup. You can even watch this decision process in a built-in web view, making it a fascinating look at how we can improve AI quality right now.

🆔 @hackernewsgithubprojects

📦 physiclaw/physiclaw

PhysiClaw: The AI That Uses a Real Hand

PhysiClaw is the AI agent that treats your phone screen like a physical surface, using a robotic arm to tap and swipe with a real stylus. It solves the problem of apps that refuse to share data or block automated tools. By watching the screen with a camera and acting like a human finger, it can handle everyday tasks like ordering food or booking rides without needing any special software installed on the device itself. It works in any app because it mimics natural touch. The result is an assistant that actually does things in the physical world, not just talks about them.

📰 https://news.ycombinator.com/item?id=49345058

🆔 @hackernewsgithubprojects

📦 figranium/figranium

Figranium: Your Own Browser Automation API

Figranium is the self-hosted tool that turns visual browser automations into instant API endpoints. Most cloud scrapers charge per click and keep your data on their servers, but this project lets you build stealthy web tasks by dragging blocks together on your own machine. It uses a unique engine to mimic human typing and mouse movements so your scripts don't get blocked. Once a task is built, you can trigger it from any app using a simple API call. This gives you full control over your data and infrastructure without paying recurring cloud fees. It is the perfect way to bring reliable, private browser automation into your own workflow.

📰 https://news.ycombinator.com/item?id=49345197

🆔 @hackernewsgithubprojects

📦 romainsimon/uisfx

UI SFX: 936 sounds, zero dependencies

UI SFX is a tiny, zero-dependency sound library that lets you completely change the personality of your app's interface sounds without touching a single line of your logic. Instead of hunting for random audio files, you call semantic cues like 'success' or 'warning', and the system plays the right sound from one of twelve distinct sonic packs. The most surprising part is that you can switch the entire vibe from 'minimal' to 'arcade' in one line of code, while the underlying interaction logic remains exactly the same.

🆔 @hackernewsgithubprojects

📦 thinkany-ai/dscode

DSCode: The Local AI Coder

DSCode is the local coding agent that finally puts you in the driver's seat. It runs right on your machine, keeping all your code and session data private instead of sending it off to some cloud server. The coolest part is its sandboxing. Every command the AI tries to run happens inside a locked-down environment, so even if the model gets a little wild, it cannot touch your actual files or network. You get the speed and low cost of DeepSeek by default, but you can easily swap in other models like OpenAI or Anthropic whenever you need them.

🆔 @hackernewsgithubprojects

📦 r4dl/come

CoMe: The Method That Turns Clouds into Solid Surfaces

CoMe is the mesh extraction tool that finally gives your 3D Gaussian Splatting scenes a solid, usable surface. You know how those new 3D camera techniques create beautiful, realistic images from thousands of tiny, fuzzy splats? The problem is those splats are just a cloud; they are not solid objects you can touch or use in a movie. CoMe solves this by using a smart confidence system to figure out exactly where the solid surface ends and the empty space begins. It is like a digital sculptor that carefully chisels away the noise to reveal a clean, detailed 3D model.

🆔 @hackernewsgithubprojects

📦 internlm/intern-s2-mobius

Intern-S2-Mobius: Faster AI Reasoning

Intern-S2-Mobius is the foundation model that separates memory from thinking to deliver speed. While standard AI systems store facts and process logic layer by layer, this thirty-five billion parameter model uses a shared memory bank that multiple reasoning units query in parallel. This design lets the system refine its thoughts internally rather than writing out long, visible step-by-step notes. The result is a massive four times speedup in inference time because the model thinks more efficiently in the background.

🆔 @hackernewsgithubprojects

📦 vixhal-baraiya/microgpt-c

MicroGPT C: Training a GPT in Pure C

Generate new names by training a character level transformer on a single line of text. MicroGPT C fits an entire neural network into one C file with no external libraries. It uses a tiny model to learn patterns and predict the next letter. The code relies only on your operating system core to handle math and memory. This project proves you can build and run a language model from scratch. You do not need heavy frameworks. You just need a compiler and a text file. The result is a clear look at how AI models actually work under the hood. It is a perfect way to learn the fundamentals.

📰 https://news.ycombinator.com/item?id=49347477

🆔 @hackernewsgithubprojects