📦 xcjy8bao/baoer_signal_grep

Baoer Signal Grep: The Librarian for AI Code

Stop letting your AI coding agent drown in messy file dumps. Baoer Signal Grep acts like a patient librarian, guiding your agent to find exact text passages without opening every single folder in your project. It uses a stable cursor system to keep your place, so you can easily pick up where you left off without losing context. The coolest part? It handles fuzzy searches by combining exact keyword matches with a local semantic model, so even if you remember a sentence slightly wrong, it still finds the right spot. No more guessing which file has the answer you need.

🆔 @hackernewsgithubprojects

📦 zenstory-ai/oh-story-dsh

Oh Story: From Novel to Short Drama

Turn raw novel text into fully produced short dramas without writing a single line of code. This repository is a specialized toolkit that automates the entire pipeline from analyzing a story to generating actual video prompts and sound design. It handles the heavy lifting of character consistency, shot composition, and even specific visual styles like ink-wash or 3D animation. You feed in a book, and it breaks down the plot into professional screenplays and technical instructions for AI media generators. It is essentially a digital film crew that ensures your story looks coherent on screen.

🆔 @hackernewsgithubprojects

📦 x-square-robot/x2streaming-tts

X2Streaming-TTS: Speak Before The Text Is Done

X2Streaming-TTS is the text-to-speech engine that starts speaking before the sentence is even finished. Usually, AI waits for a complete sentence to sound natural, which creates noticeable delays in real-time conversations. This project solves that by listening to the text as it arrives and making permanent decisions about pronunciation immediately. It never takes a word back, so if it says a number wrong, it stays wrong. To make this safe, it uses a clever trick called causal commitment. It waits just long enough to know if a number is a quantity or a rank, then locks in the correct sound.

🆔 @hackernewsgithubprojects

📦 tile-ai/tilelang

TileLang: Write Python, Get GPU Speed

TileLang is the programming language that lets you write GPU code in Python. It solves the headache of writing fast, low-level hardware kernels by letting you use familiar Python syntax, while a compiler underneath automatically handles the complex, low-level optimizations. It is perfect for building high-speed matrix math and attention operations without needing to manually manage every hardware detail. Just describe your logic in plain Python, and it compiles directly to top-tier performance on Nvidia, AMD, and Apple hardware. It is a genuinely clever way to get serious speed without the usual hardware headache.

🆔 @hackernewsgithubprojects

📦 yyjeqhc/webcodex

How WebCodex Gives AI Agents Your Real Dev Environment

WebCodex is the bridge that lets AI agents code on your actual machine instead of a cloud sandbox. Most AI coding tools require you to upload your project, but WebCodex flips that script. It runs locally on your computer and connects securely to chat assistants like ChatGPT or Claude. This means the agent can read your files, run your real tests, and use your specific Git setup without ever copying your data to a third-party server. Your code stays exactly where it is, while the AI gets full access to your local toolchain. It is a powerful way to keep your development environment private and intact.

🆔 @hackernewsgithubprojects

📦 edge0-ai/edge0

edge0

edge0 is the inference framework that runs massive language models directly from your solid-state drive without needing massive amounts of random access memory. Instead of loading the entire model into your computer's fast memory, it streams specific parts from the drive only when needed, which keeps your device running cool and efficient. It achieves this by predicting which parts are needed next and loading them in the background so the computer never has to wait. This means you can run powerful AI locally on standard consumer hardware like a Mac mini, making advanced technology accessible to anyone without a data center.

🆔 @hackernewsgithubprojects

📦 open-science-ledger/how-to-train-your-slop-cannon

How to Train Your Slop Cannon: LLM Guide for Scientists

Turn your chatbot into a rigorous research partner by fixing how scientists actually use large language models. The project called How to Train Your Slop Cannon tackles the messy reality that AI writes plausible nonsense at an industrial scale. This short guide stops you from blindly trusting the text and shows you how to verify everything. The coolest part is using structured proof formats to break big claims into tiny, checkable sub-claims. This stops the model from just hallucinating answers and forces it to show its work. It also explains why your model keeps forgetting context or agreeing with you just to be nice.

📰 https://news.ycombinator.com/item?id=49600045

🆔 @hackernewsgithubprojects

📦 alayalab/pwm

pwm

Pwm is the game engine that finally makes video generation actually playable. Most AI video models just guess what looks right, but this project separates the rules from the picture. It keeps a clear list of what is actually happening, like where your character is, and then shows that to you. That means things stay true even when you look away. You can change the world, check it later, and trust it. It is a real engine, not just a fancy filter. It is early, but the idea is huge. If you want AI that remembers and obeys rules, this is the one to watch.

🆔 @hackernewsgithubprojects

📦 foadsf/vintage-latex

Vintage Latex: The Old Science Aesthetic

Make your documents look like they were printed in the 1800s. Vintage Latex is a collection of twenty standalone examples that recreate the look of old scientific papers, complete with engraved plates, star charts, and calculated log tables. It solves the problem of modern documents feeling too digital and sterile. By using specific fonts and drawing tools, it gives your work a warm, historical weight that feels much more credible. This is a great resource if you want to add some real character to your next project. So grab an example, compile it, and see what the old masters did right.

📰 https://news.ycombinator.com/item?id=49629624

🆔 @hackernewsgithubprojects

📦 zjunlp/mechanist

Mechanist: The AI Researcher That Proves Itself

Run a full scientific investigation into how large language models actually think, without writing a single line of code. Mechanist is an autonomous agent that takes a vague research question and turns it into a verified, evidence-backed discovery. It hunts through thousands of academic papers, formulates a specific hypothesis, writes and executes the necessary experiments on your hardware, and then cross-checks its own conclusions with an independent model to ensure the results are solid. It handles the entire messy workflow from literature review to final verification, acting like a tireless lab assistant that never gets tired of checking its own work.

🆔 @hackernewsgithubprojects

📦 ucsandman/discovery-loop

How Discovery Loop Broke World Records for $28

Discovery Loop recently broke ten world records in the circle packing problem for under twenty-eight dollars. It is a local research lab that uses your existing AI subscriptions to evolve optimization solvers. Instead of trusting the models blindly, it runs them in isolated environments, demands independent mathematical checks, and requires human approval before anything counts. This rigorous loop found better solutions than previous human experts for specific packing benchmarks. It proves you can run serious scientific research from your own machine without massive API bills. If you want to build AI systems that actually verify their work, this project is the blueprint.

🆔 @hackernewsgithubprojects

📦 alibabaresearch/cut-craft-bench

CutCraft: The Video Editing Benchmark

Current AI video generators look cinematic, but they completely ignore how real film editors actually cut scenes. CutCraft is the first benchmark built specifically to test if these models understand professional editing techniques like J-cuts and transition timing. It stops relying on simple visual checks and instead measures if a model can execute specific editorial instructions, like keeping audio and video cuts in sync or building a proper montage. The project also includes an agentic baseline that breaks generation down into planning, shot synthesis, and composition to actually realize these edits.

🆔 @hackernewsgithubprojects

📦 roy-tong/agentmeasure

AgentMeasure: The Truth About AI Agent Metrics

AgentMeasure treats missing data as a valid answer rather than a hidden zero. As artificial intelligence agents start handling real work, their success rates often look inflated because a single failed task might be counted as two separate attempts. This open source standard fixes that by distinguishing between one logical goal and the multiple technical tries required to finish it. It helps you see exactly where your software fails and retries, so you can trust your numbers before customers pay for the outcome. This project turns vague activity into clear, verifiable facts.

🆔 @hackernewsgithubprojects

📦 penpot/penpot-ai-kit

Penpot AI Kit: Let AI Design Real Files

Penpot AI Kit lets you tell an AI assistant to build real, editable designs directly inside your Penpot file, rather than just generating flat images. You simply ask for a dashboard or a button component, and the kit guides the assistant to use your existing colors and shapes. The coolest part is that the AI grades its own work for quality and accessibility before showing you the result. It pauses to ask permission before making big changes, so you stay in control. This is perfect for designers who want to speed up their workflow without losing the polish of a human touch.

🆔 @hackernewsgithubprojects

📦 blacktwist/social-media-skills

Social Media Skills: Your AI Content Brain

Social Media Skills is the AI toolkit that turns your creative chaos into a clear, data-backed plan. Instead of guessing what posts will land, this collection of instructions teaches your AI assistant to analyze your past performance and identify exactly which topics and formats are driving real growth. It helps you see the difference between content that just gets likes and content that actually builds an audience. By connecting your creative output to hard follower data, it removes the guesswork from your content strategy. You get a personalized roadmap based on your specific numbers, making every post you write a calculated step toward your goals.

🆔 @hackernewsgithubprojects

📦 iml1s/resume-skills

Resume Skills: Your Coding Agent Memory Bank

Your coding assistant can forget its own work history, but resume skills fixes this by reading local conversation logs and packaging them into a safe, offline summary. This tool copies that context into a brand-new session so your agent remembers exactly what you were building. It works entirely on your computer without sending data to the cloud or restarting your old chat. The biggest benefit is that it supports almost every major coding assistant, letting you switch between tools without losing your place. It is not a magic save file that reboots a dead process, but a clean handoff of notes.

🆔 @hackernewsgithubprojects

📦 databricks-solutions/consort

Consort: The Music Ensemble For AI Code

Stop letting your AI code generator guess. Consort fixes this by acting like a music ensemble where every agent plays one specific instrument. It uses a strict conductor to enforce real tests against live database branches, so the code is actually correct. This means you get genuine engineering discipline instead of just lucky prompts. The result is a clean, verified application backend. Trust the music, not the model.

🆔 @hackernewsgithubprojects

📦 rohasnagpal/legal-ai-skills

Rohas Legal AI: 160+ Skills in One Click

Rohas Legal AI is the all-in-one toolkit that turns any chatbot into a legal assistant. It bundles over one hundred and sixty specific skills, covering everything from drafting contracts to analyzing arbitration awards, all wrapped in a single easy-to-install plugin. The coolest part is how it works: you simply describe your legal task in plain everyday language, and the system automatically picks the right skill from the library to handle it. You never have to remember a specific command or search for the right tool. It is built for lawyers, law students, and business owners who want accurate, structured legal help without the complexity of managing individual prompts.

🆔 @hackernewsgithubprojects

📦 quome-cloud/openworld

openworld

Building a working model of a complex system usually requires massive datasets, but this project claims you can do it in under two minutes. Openworld functions like a universal toolkit for creating these internal simulations, allowing you to test how code behaves before you ever deploy it. It solves the problem of needing expensive training data by constructing a logical map of the environment from the rules alone. You essentially build a safe, digital playground where you can run your software through hundreds of different scenarios instantly. This is useful because it turns guesswork into a verifiable process, letting you catch errors before they break your production systems.

🆔 @hackernewsgithubprojects

📦 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