📦 openlair/openskill

Stop Training AI: Let Agents Learn for Themselves with OpenSkill

Imagine if AI agents could teach themselves new skills without human help. That is exactly what OpenSkill solves. Traditional AI requires hand-curated data, but OpenSkill lets agents build both their own skills and verification signals from scratch using only a task prompt and open-world resources. By retrieving knowledge from the web and refining skills through self-built virtual tests, agents can evolve in total isolation from target-task supervision. This new paradigm makes AI surprisingly capable and grounded, achieving performance that rivals human-level benchmarks. Stay tuned as this framework evolves, because the future of autonomous agent learning is just getting started.

🆔 @hackernewsgithubprojects

📦 dataarctech/bayesian-agent

Stop Prompting, Start Evolving: Meet Bayesian-Agent

Tired of manual prompt engineering? Discover Bayesian-Agent, a powerful framework that turns your AI agent trajectories into reusable, evidence-weighted skills and procedures. Instead of guessing what works, it uses a Bayesian layer to treat every success and failure as data, allowing agents to evolve their own operational knowledge across different platforms. Whether you are repairing failed tasks incrementally or building complex workflows from scratch, this tool helps you optimize performance with actual verified evidence. It even features a native harness for running your own LLM loops. Start treating your agent's experiences as assets and level up your automation strategy today.

🆔 @hackernewsgithubprojects

📦 dmmm1997/videoseg-o3

VideoSEG-O3: Mastering Video Object Segmentation with Reinforcement Learning

Ever wonder how AI can track objects in videos with such precision? Meet VideoSEG-O3, a cutting-edge reinforcement learning framework designed to master reasoning video object segmentation. Instead of just looking at fixed frames, this system actively explores temporal intervals and keyframes to understand object identity, motion, and complex linguistic cues. By using a decoupled thinking trace, it structures reasoning into a clear multi-turn workflow, aligning model decisions with high-quality pixel-level masks through advanced reinforcement learning. Whether you are tackling intricate visual tasks or pushing the boundaries of spatial-temporal AI, VideoSEG-O3 provides a smarter way to see the world.

🆔 @hackernewsgithubprojects

📦 centrechen/embfilter

Optimize Your Text Embeddings with EmbFilter

Want to boost your text embeddings without the heavy lifting? Check out EmbFilter, a lightweight linear filter designed to refine zero-shot text embeddings effectively. Based on the research paper exploring how embedding matrices act as a feature lens, this tool allows you to compress and clean up your data with ease. You can simply specify a filter ratio to determine how many dimensions to save, giving you better control over your model performance. It is a smart, streamlined way to enhance your text processing pipeline. Dive into the repository to see how it can sharpen your projects today.

🆔 @hackernewsgithubprojects

📦 marinero4972/awesome-humanview-videounderstanding

Mastering Video AI: A Human-View Approach

Ever wonder how artificial intelligence actually makes sense of the hours of video footage it processes? This repository features a comprehensive survey on human-view video understanding using large multimodal models. It breaks down complex video analysis into three cognitive pillars: watching for perceptual grounding, remembering to maintain context over long streams, and reasoning to derive evidence-based answers. Whether you are interested in fine-grained temporal grounding or efficient long-video processing, this collection provides a structured look at the latest methods and benchmarks in the field. Dive into this academic survey to see how researchers are building smarter, more observant video intelligence.

🆔 @hackernewsgithubprojects

📦 melandlabs/openloomi

Stop Forgetting! Meet OpenLoomi: Your Proactive AI Memory

Are you tired of losing track of important work details across endless apps? OpenLoomi is an open-source, local-first AI workspace that acts as a proactive partner for your professional life. It connects directly to your favorite tools like Slack, Jira, email, and Google Drive to build a comprehensive context graph of your projects, decisions, and communications. By running locally with AES-256 encryption, it keeps your data secure while providing intelligent, context-aware assistance that anticipates your needs. Whether it is tracking documents or automating tasks, OpenLoomi ensures your digital workspace finally works for you. Take control of your workflow and stay focused today.

📰 https://news.ycombinator.com/item?id=48460969

🆔 @hackernewsgithubprojects

📦 fareedkhan-dev/all-agentic-architectures

Master 35 Agentic AI Architectures in One Library

Stop guessing which AI agent pattern actually works and start building with a professional standard. This repository is a powerful Python library and runnable textbook that puts thirty-five production-grade agentic architectures, like Reflexion, GraphRAG, and MemGPT, directly into your hands. It solves the chaos of choosing the right approach by providing a unified interface and a comparative benchmark leaderboard for seventeen distinct tasks. With multi-provider support, you can easily swap models while keeping your logic intact. Dive into these live, executed examples to sharpen your skills and finally master the complex world of agentic AI design.

🆔 @hackernewsgithubprojects

📦 giovannipasq/agentic-rag-for-dummies

Build an Agentic RAG System in Minutes

Want to build an intelligent search system that actually understands context? This repository provides a modular blueprint for an Agentic RAG system using LangGraph. It solves the typical RAG limitations by utilizing hierarchical indexing and multi-agent reasoning, which splits complex questions into manageable sub-queries for higher accuracy. You get professional features like conversation memory, automatic query clarification, and even a human-in-the-loop workflow. It is designed to be provider-agnostic, letting you switch between major LLMs with ease. Whether you are learning core concepts or building a scalable production-ready architecture, this project offers the perfect foundation to start your agentic AI journey today.

🆔 @hackernewsgithubprojects

📦 vstorm-co/pydantic-deepagents

Build Autonomous AI Agents with Pydantic Deep Agents

Ever wonder how to build production-grade autonomous agents that actually get things done? Meet the harness that transforms any language model into a capable, functional assistant. Instead of starting from scratch, you get a complete infrastructure for planning, memory, and sandboxed code execution right out of the box. Whether you need to automate complex file operations, delegate tasks to subagents, or maintain infinite context through smart summarization, this framework handles the heavy lifting. With built-in support for Docker workspaces and custom skill definitions, you can finally move from simple chat to reliable, multi-step automation. Take control of your AI workflow today.

🆔 @hackernewsgithubprojects

📦 souravroy-etl/duckle

Duckle: The Local-First AI ETL Studio You Need to Try

Tired of complex server setups just to move your data around? Meet Duckle, an open-source, local-first ETL studio that brings professional data engineering right to your laptop. With its drag-and-drop canvas, you can easily design pipelines and execute them at native speeds using DuckDB. The best part is the built-in AI assistant, Duckie, which runs entirely on your CPU to write your pipeline logic in plain English. With over two hundred connectors and built-in Git support, it is the perfect tool for developers who value privacy and speed. Download this tiny, powerful app and start building data pipelines locally today.

🆔 @hackernewsgithubprojects

📦 denoland/clawpatrol

Stop Rogue AI Agents: Meet Claw Patrol

Are you concerned about AI agents accessing sensitive production environments? Meet Claw Patrol, a security firewall designed specifically to gate agent actions. By sitting directly between your agents and your infrastructure, it parses traffic at the wire level and enforces granular security policies written in simple configuration files. For example, you can block destructive database commands or require human approval before specific actions like deleting Kubernetes resources occur. It provides a secure tunnel for your agents, ensuring every request is safe. Take control of your agent traffic today and prevent unauthorized access with this powerful, unified security gateway.

📰 https://news.ycombinator.com/item?id=48462928

🆔 @hackernewsgithubprojects

📦 yv1ing/z3r0

Automate Your Red-Teaming with AI Agents

Ever wonder how to streamline complex penetration testing? Meet Z3r0, an AI-native red-team workbench designed for authorized security research. Instead of manual grunt work, it coordinates a team of specialist agents for tasks like reconnaissance, code auditing, and reverse engineering. It solves the chaos of managing long-term security engagements by using sandboxed Docker environments for tools and preserving findings as durable evidence records. With its unique timeline replay and structured attack-path analysis, you can audit every step of your assessment with confidence. Take control of your security workflow and see how automated, agent-driven operations can transform your penetration testing process today.

📰 https://news.ycombinator.com/item?id=48462840

🆔 @hackernewsgithubprojects

📦 apple/coreai-models

Build On-Device AI with Apple's Core AI Models

Ever wonder how to get powerful machine learning models running directly on your iPhone or Mac? The Core AI Models repository is your go-to toolkit for exactly that. It provides essential export recipes to convert popular open-source models into the optimized format needed for on-device performance. You get access to Python primitives for building custom models in PyTorch, plus handy Swift runtime utilities to integrate them into your apps. Whether you are working with language, vision, or audio models, this project simplifies the path to efficient local AI. Start exploring the catalog today and bring smarter features to life.

🆔 @hackernewsgithubprojects

📦 ratel-ai/ratel

Stop Stuffing Your AI Agent with Too Many Tools

Are you wasting tokens by stuffing your AI agents with every available tool? Meet Ratel, an in-process context engineering platform designed to fix that. Instead of overloading your model, Ratel acts as a smart filter between your agent and its catalog. Using deterministic BM25 retrieval, it intelligently selects only the most relevant tools for each turn. It is a lightweight, framework-agnostic library that requires no vector databases or complex infrastructure. By keeping your context clean, you can significantly reduce costs and improve performance. Start optimizing your agent’s context window today and keep your AI focused on what actually matters.

🆔 @hackernewsgithubprojects

📦 hongsong-wang/propdiff-tmg

AI Designing Materials: Property-Informed Microstructure Generation

Ever wonder how AI can help invent new materials? This research project introduces a cutting-edge diffusion model designed to generate complex 3D voxel microstructures from simple text descriptions. It goes beyond aesthetics by incorporating physical property controls, allowing you to specify metrics like volume fraction, elastic modulus, and Poisson's ratio during the generation process. By combining deep learning with structural engineering, this tool offers researchers a powerful way to create precise, performance-optimized materials automatically. Whether you are in materials science or generative AI, this innovative approach is transforming how we engineer the microscopic building blocks of our future world.

🆔 @hackernewsgithubprojects

📦 wbopan/retro-harness

Supercharge Your AI Agents Without Human Labels

Want to make your AI agents smarter without needing expensive human-labeled data? Meet Retrospective Harness Optimization, or RHO. Instead of relying on external feedback, RHO allows agents to improve their performance by analyzing their own past experiences. It intelligently selects diverse tasks, performs parallel rollouts, and uses self-consistency checks to refine its own skills and tools. In practice, this approach can dramatically boost success rates on challenging benchmarks by learning directly from unlabeled trajectories. It is a powerful way to automate agent improvement, so start evolving your agent's capabilities today and see what it can achieve.

🆔 @hackernewsgithubprojects

📦 huixiai/rhinovla

RhinoVLA: Real-Time Robot Control on Edge AI Chips

Imagine robots that think and react in real-time, all running directly on the hardware. RhinoVLA is a groundbreaking cross-embodiment system designed for edge-side robot control. By combining a powerful model with the efficiency of the R1 chip, it achieves an impressive inference speed that meets the gold standard for closed-loop control. It uses clever techniques like instance adaptation to work seamlessly across different robot platforms, from instruction following to complex bimanual tasks. This project bridges the gap between sophisticated artificial intelligence and physical robotics, proving that real-time, high-performance edge computing is the future of intelligent autonomous systems.

🆔 @hackernewsgithubprojects

📦 5uck1ess/tts-bench

Stop Guessing: How to Benchmark Your AI Voice Models

Are you struggling to find the best local text to speech model for your specific hardware? This repository provides a comprehensive benchmarking suite designed to evaluate speed, quality, and voice cloning capabilities across diverse platforms including Windows, Linux, and Mac. By measuring cold and warm latency, memory usage, and real-time factors, it helps you identify the most efficient models for your system. It even includes a blind voting arena where you can compare audio clips to find the most natural sounding voices. Stop guessing and start testing your models today to get the performance and quality you truly need.

🆔 @hackernewsgithubprojects

📦 luke8086/gentleos32

Build Your Own OS for Vintage PCs

Ever wanted to breathe new life into an old 32-bit PC? Meet GentleOS, a minimalist hobby operating system designed specifically for vintage hardware. This project provides a clean, monolithic platform that runs directly on bare metal, requiring only an i386 processor and a little RAM. It handles core essentials like VGA graphics, keyboard input, and even a functional mouse interface to support interactive graphical apps. Whether you are tinkering with retro setups or just curious about kernel development, this open-source project offers a fascinating, hands-on way to explore how operating systems talk to hardware. Dive in and start building today.

📰 https://news.ycombinator.com/item?id=48458890

🆔 @hackernewsgithubprojects

📦 cporter202/api-mega-list

The Ultimate API Treasure Trove for Your Next Project

Are you tired of hunting for the right tools to fuel your latest app idea? This repository is a goldmine for developers and creators, offering a massive collection of ready-to-use APIs designed to power everything from simple automations to complex applications. Whether you need to extract YouTube transcripts, automate job data collection from platforms like LinkedIn, or integrate advanced AI assistants using the Model Context Protocol, this list has you covered. It solves the headache of searching for reliable services by gathering high-quality, practical tools in one place. Explore this library today to streamline your workflow and start building faster.

🆔 @hackernewsgithubprojects