📦 veris-ai/riley-agent

Riley Agent: Test Voice AI with Real Banking Data

Riley Agent is the open-source benchmark that finally lets you compare every major voice AI platform using the same realistic banking support scenario. It solves the messy problem of testing voice assistants by giving you a consistent, simulated customer service workflow where an AI agent named Riley helps users manage fake credit cards. Instead of wrestling with different setups, you can drop in implementations from giants like OpenAI, Google, or AssemblyAI and watch them handle the exact same conversations with synthetic bank data.

🆔 @hackernewsgithubprojects

📦 gen-verse/past-bench

How Agents Actually Learn From Mistakes

PAST-Bench tests if personal AI agents can genuinely improve over time by learning from past mistakes. Instead of grading isolated tasks, it runs agents through long sequences where early experiences are supposed to help them handle later challenges better. The repo includes a clever framework called Hermes+ that adds specific memory and routing tools to help agents remember and reuse what they have learned. It essentially asks whether an AI remembers its previous failures and uses that knowledge to act smarter next time, rather than starting from scratch every time you talk to it.

🆔 @hackernewsgithubprojects

📦 msp-xen/ae-uav

AE-UAV: The First Air-to-Air Event Camera Drone Dataset

AE-UAV is the first air-to-air event camera drone dataset that turns jagged tracking labels into smooth, continuous motion paths. Most video datasets only check if an object is inside a box every few frames, but AE-UAV captures billions of raw sensor events and maps a perfect, curve-fitted trajectory across every single microsecond. This means researchers can evaluate drone tracking systems at any speed without guessing what happened between frames. It features nearly two hundred flight sequences captured from one drone chasing another, complete with thermal and RGB video for extra context.

🆔 @hackernewsgithubprojects

📦 aweai-team/calibforge

CalibForge: AI Terminal Calibration

CalibForge transforms how artificial intelligence learns to use command lines by filtering out tasks that are either too easy or impossible to solve, leaving only the perfect middle ground for training. Instead of dumping thousands of random coding challenges on a model, this project uses an adversarial process where different AI agents try each task and only the ones with mixed results survive. This creates a curated dataset of five thousand precise terminal exercises that force the learner to actually think, rather than guess. By training on this calibrated data, models see massive performance jumps on software engineering benchmarks because they are learning from genuine difficulty.

🆔 @hackernewsgithubprojects

📦 fudan-generative-vision/dynetrion

Dynetrion: The AI That Simulates Protein Movement

DyneTrion is the generative model that finally captures how proteins move and change shape over time. Most AI tools only guess what a static molecule looks like, but this project simulates the actual physical dance of atoms across different time scales. It solves the headache of predicting how complex biological structures shift and flex in real life. By training on massive datasets of molecular dynamics, it creates smooth, realistic animations of these tiny machines working. This matters because seeing the motion helps scientists understand diseases and design better medicines. It turns confusing data into clear, playable video clips of biology in action.

🆔 @hackernewsgithubprojects

📦 mbbill/mind-expander

Mind Expander

AI coding tools write fast code but explain it through exhausting walls of text. Mind Expander changes that by instantly turning your Rust or TypeScript codebase into an interactive, infinite-canvas browser diagram. Instead of reading long markdown summaries, you and your AI agent work on the exact same visual graph of modules, dependencies, and refactoring plans. It translates complex relationships and system architecture into a clear, source-backed visual tour you can click through. It is like having a GPS map for your system architecture that actually links back to your live files, making code reviews and refactoring plans incredibly simple to verify.

📰 https://news.ycombinator.com/item?id=48281337

🆔 @hackernewsgithubprojects

📦 derec4/internships-and-newgrad

Internships and New Grad

Save hours of searching by using a tool that automatically aggregates and deduplicates tech internship and new grad positions into one clean list. This project runs a Go program that fetches data from multiple popular job boards, parses the results, and merges them to remove duplicates while sorting by posting date. It creates a master spreadsheet you can actually use, turning scattered listings into a single source of truth for your job hunt. It is a practical way to stay ahead without manually copying and pasting from ten different sites. Grab the code, run it, and let the automation handle the busy work for you.

🆔 @hackernewsgithubprojects

📦 oil-oil/oil-motion

How to Make AI Video Respond to Your Scroll

Oil Motion turns AI video from a passive clip into an interactive tool that follows your scroll, mouse, or touch. It does this by first locking down key start and end frames to keep the subject consistent, then using AI to fill in the smooth motion between them. The system cleans up the video by removing awkward pauses and compressing it into efficient web assets. Instead of just playing from start to finish, the code maps your physical actions—like dragging a slider or scrolling down the page—directly to specific frames in that animation. This means you can pause, rewind, or scrub through the video naturally while watching it.

🆔 @hackernewsgithubprojects

📦 hakanseven12/opencadstudio

Open CAD Studio: Free 2D and 3D CAD in Rust

Open CAD Studio lets you draft precise 2D blueprints and model 3D shapes right in your browser. Built in Rust, it handles heavy files like DWG and DXF without choking, thanks to GPU-accelerated rendering that makes the interface feel snappy even with complex designs. It even lets you pull in external reference drawings so multiple files stay in sync. The best part? It runs locally or online, keeping your data private while giving you professional tools for free. If you need reliable CAD that respects your privacy and doesn't cost a dime, this is the real deal.

🆔 @hackernewsgithubprojects

📦 jordanrendric/claude-video-vision

Claude Code Video Vision

Watch Claude actually see and hear videos by passing a YouTube link or local file into Claude Code and getting back clear frame images and spoken words. This plugin uses ffmpeg to pull key pictures and sends the sound to your preferred audio tool, whether that is the Gemini API, OpenAI, or a fully local whisper setup on your own machine. It is genuinely clever because Claude adjusts the video quality and frame rate based on what you ask, so you get detailed shots for specific moments and lighter previews for long summaries.

🆔 @hackernewsgithubprojects

📦 universal-commerce-protocol/ucp

Universal Commerce Protocol Explained

UCP is the open standard that lets AI agents and apps talk to stores without messy custom code. Instead of building one-off integrations for every shop, businesses publish a simple profile declaring what they can do, like checkout or loyalty rewards. Your favorite app or bot reads that profile, discovers the capabilities, and handles the rest using secure, standard methods. It means your smart assistant can find products, manage your cart, and pay for you across any site that supports the protocol. No more manual data entry or risky workarounds.

🆔 @hackernewsgithubprojects

📦 minimax-ai/minimax-h3

MiniMax H3

MiniMax H3 is the open-source video model that finally brings professional stereo sound and high-resolution visuals to your local machine. While most AI video tools are stuck producing silent clips or blurry previews, this project gives you a 33-billion-parameter engine that generates crisp 2K video paired with native stereo audio tracks. It is genuinely interesting because it treats sound as a first-class citizen rather than an afterthought, using a unified transformer to predict both visuals and audio simultaneously. You can feed it simple text prompts, still images, or even reference videos, and it will spin up fifteen-second clips with natural-sounding soundscapes.

🆔 @hackernewsgithubprojects

📦 nicobailon/pi-intercom

How AI Agents Talk to Each Other on Your Computer

Ever wonder how to make multiple AI coding sessions actually collaborate in real time? Pi Intercom lets you set up direct one-to-one messaging between different AI sessions running on the same machine. Instead of a messy group chat, you can pick a specific session and send it context, questions, or tasks. One session can research a problem, then ask another session for help, and even wait for the answer before moving forward. It is perfect for splitting work between a planner and a worker, or for a main agent to check in with its helper tools.

🆔 @hackernewsgithubprojects

📦 humza-khalid/12vhpwr-guard

12vhpwr Guard

This project protects your expensive graphics card by monitoring the electrical current flowing into its power connector and automatically shutting down your computer if it detects dangerous levels. The 12vhpwr Guard tool runs quietly in the background on Windows, constantly checking the specific current readings from your GPU’s internal sensors. If it spots a sustained overload that could melt the connector, it forces an immediate system shutdown to prevent permanent hardware damage. It works specifically with ASUS ROG Astral cards, using direct sensor access to avoid needing extra software. Think of it as a dedicated safety net that watches over your high-power components when standard protections might fail.

🆔 @hackernewsgithubprojects

📦 duckyshell/comfyui-minimaxh3-firstblockcache

Speed Up Video Generation with MiniMax H3 FirstBlockCache

Speed up your ComfyUI video generation by skipping redundant math steps, thanks to a clever node called MiniMax H3 FirstBlockCache. This tool sits between your model and the scheduler, running the first processing block normally but then caching the result for the next step. If the video hasn't changed much, it skips the heavy lifting entirely, giving you a thirty percent boost in speed without needing special hardware. You get three smart modes to balance speed and quality, letting you choose how much you trust that cached math.

🆔 @hackernewsgithubprojects

📦 tanstack/charts

Charts

TanStack Charts lets you write a single chart definition that instantly works across React, Vue, Angular, and even runs on the server without any browser. The biggest surprise is that you do not need to rewrite your code to switch between standard SVG rendering and a high-performance Canvas engine; you just change one import statement. This means you get accessibility and search engine friendly output by default, but you can still handle massive datasets by swapping the renderer without touching your chart logic. It keeps your data format simple and uses familiar tools under the hood so you stop juggling multiple libraries.

🆔 @hackernewsgithubprojects

📦 saganaki22/comfyui-sol-attn

ComfyUI Sol-Attn: Faster Video with Lower VRAM

ComfyUI Sol-Attn slashes peak video memory usage by thirty-seven percent while running noticeably faster than standard attention methods. This ComfyUI plugin uses a specialized Triton kernel to skip unnecessary memory copies and smartly approximate complex attention patterns. It is designed for modern NVIDIA Blackwell GPUs but also works on older RTX 40-series cards. The project offers specific nodes that let you ramp sparsity over time and chunk processing to keep your GPU from choking on long sequences. By reducing peak VRAM and speeding up generation, it makes running high-resolution video models much more practical on consumer hardware.

🆔 @hackernewsgithubprojects

📦 halcyon-video/halcyon-video

Halcyon Video: Walk Through Your Movie Collection

Halcyon Video turns your personal media library into a walkable 1990s video rental store that you can explore in first person. Built with three.js, it connects to your Jellyfin server and renders every film as a physical case on a shelf. You walk the aisles, pull a title off the wall, and inspect the back of the box for real reviews and technical specs. You carry the tape to the counter where a cloth simulation bag crinkles around it, then head home to a virtual living room to watch it on a CRT.

📰 https://news.ycombinator.com/item?id=49211158

🆔 @hackernewsgithubprojects

📦 smol-env/smol

smol: The Tiny AI Agent

smol lets you run a tiny AI agent that actually talks to the shell. Instead of bloated frameworks, this project gives you a minimal Python or Go script that connects to any OpenAI compatible endpoint and executes commands directly. It strips away all the fluff, system prompts, and heavy dependencies so you get a pure loop of asking and doing. The coolest part is how it trusts the model to run shell commands without needing a complex orchestration layer. You just feed it an API URL, type your request, and watch it work.

🆔 @hackernewsgithubprojects

📦 onoja123/mobile-experiments

Mobile Experiments

Make mobile interfaces feel alive by turning boring taps into satisfying animations. Mobile-experiments is a collection of standalone React Native apps that focus entirely on that one tricky part: interaction. Instead of building entire complex apps, this repo isolates specific gestures and motion effects so you can see exactly how they work. My favorite is the bookmark dissolve, where deleting a card actually shatters it into little particles. It sounds small, but seeing that kind of playful feedback makes an app feel much more responsive and fun to use. If you want to steal these ideas or just watch the code in action, check the project out.

🆔 @hackernewsgithubprojects