📦 openmoss/moss-transcribe-diarize

moss-transcribe-diarize

The open-source model moss-transcribe-diarize beats both Google and OpenAI in transcription and speaker-tracking accuracy on challenging multi-speaker benchmarks. Instead of the usual headache of gluing together separate transcription and speaker-labeling tools, this clever system does everything at once in a single, unified pass. It takes raw, messy audio from meetings, podcasts, or videos and outputs a clean transcript with exact timestamps and clear speaker tags already baked in. You can run it locally, serve it with high-speed inference engines, or use its built-in subtitle web app to burn captions straight to video. It is a brilliant, self-contained way to make sense of long-form conversations.

🆔 @hackernewsgithubprojects

📦 cloudrift-ai/emmy

Emmy PyTorch Compiler

You can compile PyTorch models directly into raw, optimized CUDA kernels using a custom compiler stack written in just five thousand lines of Python. An open-source tool called emmy lets you trace any PyTorch module, decompose it, and fuse the operations into highly efficient hardware-native instructions. Instead of dealing with massive, opaque compiler frameworks, you get to inspect your model across six distinct stages of intermediate representation all the way down to the final CUDA code. It makes optimization completely transparent, allowing you to easily benchmark your custom kernels, run experiments, and deploy tuned models to GPU servers.

🆔 @hackernewsgithubprojects

📦 jfrog/boost

Boost

Slash the token bill of your coding assistants by wrapping your terminal commands. This command-line utility intercept output from build, test, and package tools and condenses noisy terminal logs into ultra-compact, structured context. Instead of letting thousands of lines of install logs or repetitive build layers eat up your model's context window, it preserves only the critical details your agent needs to reason, like error messages, test failures, and timing statistics. This keeps your agents fast and accurate without wasting money. You can even run a report in your browser to see exactly how many tokens you saved.

🆔 @hackernewsgithubprojects

📦 harshaneel/humanize

Humanize

Write text that reads as genuinely human-authored by applying research-backed writing tweaks directly inside your favorite AI chat agent. This clever open-source repository contains humanize, a rule-based system prompt designed to strip away unmistakable AI writing patterns. By injecting natural sentence length variety, slashing predictable transition words, and matching your personal writing style from a quick sample, it helps your drafts sound authentic. It even includes a second forensic tool called ai-check to audit your work for hidden machine-generated signals. It is a brilliant, zero-dependency way to make your AI writing sound like an actual person wrote it.

📰 https://news.ycombinator.com/item?id=48820564

🆔 @hackernewsgithubprojects

📦 zymazza/mazzap

Mazzap

Mazzap is the 3D digital-twin engine that transforms open geospatial data into an interactive, self-contained virtual land model right in your browser. Unlike traditional mapping software, it runs completely locally with zero cloud dependencies, allowing you to click anywhere to read exact GPS coordinates, simulate active wildfires, and trace water flow paths across real terrain. It combines local astronomy, viewshed calculations, and vegetation mapping into a lightweight browser viewer. You can even query the land using natural language with a built-in assistant. It is a brilliant tool for visualizing how water, fire, and sunlight interact with the physical world.

📰 https://news.ycombinator.com/item?id=48806109

🆔 @hackernewsgithubprojects

📦 pglayers/pglayers

pglayers

The PostgreSQL distribution that packages extensions as ready-to-use Docker image layers, letting you mix and match features with zero compilation. Normally, official Postgres Docker containers ship entirely empty of extensions, forcing you to manually install dependencies or compile things like geospatial tools and vector search from source. This project publishes over fifty popular extensions as isolated, pre-built Docker layers. You simply stack them onto the official Postgres base image using a single copy line per extension in your Dockerfile. It completely eliminates file conflicts and manual setups. Grab the pre-configured full bundle or build your own custom database stack in seconds.

📰 https://news.ycombinator.com/item?id=48749752

🆔 @hackernewsgithubprojects

📦 servicenow/enterpriseops-gym

EnterpriseOps-Gym

EnterpriseOps-Gym is the interactive sandbox that finally tests how AI agents handle messy, multi-step office work instead of just answering trivia. While most benchmarks use simple text files, this project builds a containerized environment mimicking real corporate infrastructure across eight departments, from HR to customer service. It throws over a thousand expert-curated tasks at your agent, forcing it to navigate hundreds of tools and databases. Instead of just checking if the AI typed the right words, it runs verification queries to see if the actual environment state changed correctly. It is a brilliant reality check proving that autonomous office assistants still have a long way to go.

🆔 @hackernewsgithubprojects

📦 aerovato/magic-compact

Magic Compact

Magic Compact stops AI coding assistants from losing their minds by changing how they compress chat histories. Instead of flattening your entire programming session into a single generic paragraph that gives the AI amnesia, this tool condenses past assistant responses into local, high-fidelity summaries while leaving your original instructions and tools completely intact. Massive files and tool outputs are pruned to save tokens, but you can instantly pull them back on demand using a dedicated retrieval tool. It is a brilliant way to slash token costs and preserve the model's memory during long development sessions. Give it a try to keep your AI context clean and sharp.

🆔 @hackernewsgithubprojects

📦 egorshest/webgl-ascii-hero

webgl-ascii-hero

You can transform any standard 3D model into an interactive piece of retro terminal art directly inside a modern web browser. A project called webgl-ascii-hero is a boilerplate for Next.js that renders 3D files and instantly processes them into animated ASCII text using custom graphics shaders. Instead of just showing a flat image, this setup lets visitors drag to rotate the model and watch the character grid dynamically recalculate in real time. It even layers on retro CRT scanlines and glow effects that react to the cursor. It is a fantastic starting point for building a highly memorable website landing page.

🆔 @hackernewsgithubprojects

📦 cyl19970726/poly-sdk

Poly-SDK

This single TypeScript SDK lets developers interact with every layer of Polymarket, from placing advanced limit orders to orchestrating automated arbitrage strategies. Instead of juggling fragmented official APIs and smart contracts, poly-sdk wraps trading services, on-chain contract actions, and live WebSocket price feeds into a unified toolkit. It even simplifies the mathematical mirror properties of prediction market orderbooks to prevent double-counting. For developers building trading bots or copy-trading platforms, this tool manages rate limits, caches active data, and handles complex split and merge operations under the hood, making it a reliable foundation for building custom trading tools.

🆔 @hackernewsgithubprojects

📦 deepfoldprotein/patchr

patchr

Most experimental molecular structures in the Protein Data Bank are missing critical parts like flexible loops or unresolved sidechains. Patchr fixes this by using diffusion-based AI to perfectly fill in the blanks while keeping your original, hard-won coordinates exactly as they are. It works across proteins, DNA, RNA, and complex multi-chain structures, boasting a ninety-nine percent chain connectivity success rate. Even better, you can go straight from fixing your molecular structure to running molecular dynamics simulations in Gromacs or Amber. It is like content-aware fill, but for the fundamental building blocks of life.

📰 https://news.ycombinator.com/item?id=48835428

🆔 @hackernewsgithubprojects

📦 chrishayduk/nanofold-competition

nanofold-competition

AI models usually get smarter by throwing mountains of data at them, but biology does not work that way because high-quality protein data is incredibly scarce. The nanofold-competition repository turns this bottleneck into a brilliant slowrun tournament where developers compete to build the most data-efficient protein folding models. Everyone gets the exact same budget of ten thousand protein structures and is challenged to see who can learn complex physical geometry fastest. By restricting training sizes and evaluating how quickly your model learns, this project forces you to design smarter biological priors instead of just renting more GPUs. It is a fantastic playground for anyone wanting to test their machine learning...

🆔 @hackernewsgithubprojects

📦 meigen-ai/genevolve

GenEvolve

Design highly accurate, complex images by turning AI image generation into a cooperative, multi-step planning process. Instead of just trying to create a picture from a simple prompt, genevolve acts like a smart project manager. It actually searches the web for real facts, gathers visual reference images, and applies specific drawing skills like anatomy and text rendering to build a detailed blueprint. This blueprint is then handed off to a separate image generator to render the final masterpiece. By breaking the work down, it solves the problem of AI getting confused by complex details, making it incredibly useful for creating stunning, context-aware visuals with actual real-world accuracy.

🆔 @hackernewsgithubprojects

📦 razee4315/paperling

Paperling

Paperling is the no-setup Markdown editor that lets you open and work with any document instantly without the overhead of heavy note-taking apps. Built with Tauri and React, it delivers lightning-fast native performance while keeping your workspace completely distraction-free. The absolute best part is its optional, local-first AI assistant: you simply describe what you want, and it proposes live, inline changes directly in the text as a color-coded comparison. You can review, accept, or reject each edit on the spot, keeping you in complete control. Grab the installer for your platform and experience friction-free writing today.

🆔 @hackernewsgithubprojects

📦 playableintelligence/game-creator

Game Creator

Game Creator is the game development toolkit that turns text prompts into fully playable browser games in minutes. Built for AI coding assistants like Claude, this smart system scaffolds two-dimensional or three-dimensional games, writes clean code, and generates custom pixel art and procedural music. What makes it brilliant is its built-in automated QA runner, which spins up a headless browser to playtest your game, catch glitches, and automatically patch the code. It is an amazing way to watch a rough game idea instantly come to life and deploy directly to the web.

🆔 @hackernewsgithubprojects

📦 justvugg/colibri

Colibri: Run a 744B AI Model on 25GB RAM

Run a massive seven-hundred-and-forty-four-billion-parameter AI model on a regular home computer with just twenty-five gigabytes of RAM. This project makes it possible by using a clever streaming trick. Instead of trying to cram the entire model into memory, it keeps the core language components in your RAM and streams the specialized neural experts directly from your solid-state drive only when they are needed. It is written in pure C with zero external dependencies. While it is not lightning fast, it proves you can run a frontier-class model on incredibly humble hardware. Check it out and see what your machine can do.

📰 https://news.ycombinator.com/item?id=48842459

🆔 @hackernewsgithubprojects

📦 linuxrebel/docubrowser

DocuBrowser

Docubrowser instantly turns a chaotic mountain of local files into a smart, searchable library that understands what your documents actually mean. By combining traditional keyword lookups with local artificial intelligence, it lets you search your PDFs, spreadsheets, and books using everyday language. You can search for concepts rather than exact phrasing, and get a custom book-jacket summary of any file on demand. Because everything runs entirely on your own computer using local AI models, you never have to worry about internet connections, subscription accounts, API keys, or your private data leaving your machine. It is a private, powerful search engine for your local hard drive.

📰 https://news.ycombinator.com/item?id=48837110

🆔 @hackernewsgithubprojects

📦 leaningtech/browsercode

Browsercode

browsercode is the browser runtime that lets you run full AI coding command line agents completely client-side. Instead of spinning up heavy cloud containers or giving automated bots risky access to your local operating system, this tool boots a secure sandbox directly in your browser. It uses WebAssembly to run Node.js, a virtual filesystem, and essential development tools like bash and git with zero installation. You can launch unmodified AI agents like Claude Code or Gemini CLI instantly, write and run web apps, and even preview the live results on the fly. It is a brilliant way to experiment with agentic coding safely.

📰 https://news.ycombinator.com/item?id=48049713

🆔 @hackernewsgithubprojects

📦 code3-dev/ccloud

ccloud

ccloud is the Android streaming app that finally brings clean, ad-free movie and TV browsing straight to your phone and television. Instead of dealing with clunky web browsers or annoying pop-ups, this clever tool lets you easily stream and download your favorite videos through a sleek, custom player. What makes it incredibly cool is how it automatically adapts to whatever device you are using, switching to a specialized sidebar layout and full remote control support when you open it on Android TV. If you want a smooth, open-source streaming hub on your own terms, you definitely need to grab this setup.

🆔 @hackernewsgithubprojects

📦 sijuec/eridani-speak

eridani-speak

Most artificial intelligence models waste your money on polite filler phrases, but eridani-speak slashes your token bill by forcing your AI assistant to communicate like a highly efficient alien engineer. Inspired by the hyper-dense language of the Eridian character Rocky from Project Hail Mary, this repository provides two clever prompt rules that strip away unnecessary words while keeping your code intact. You can choose Signal mode for ultra-compact, math-like developer notation, or Rocky mode to give your coding partner a warm but blunt personality. It is a brilliant way to cut down on response times and save token costs. Try installing it to your favorite developer environment today.

📰 https://news.ycombinator.com/item?id=47863008

🆔 @hackernewsgithubprojects

📦 binomica-labs/splicecraft

SpliceCraft

Design and simulate your entire genetic engineering workflow directly in your terminal without ever opening a browser or creating a cloud account. This clever terminal app lets you edit DNA sequences, design primers, and run virtual gel runs using crisp braille graphics right on your screen. It handles everything from simple restriction digests to complex multi-part assemblies, while keeping your valuable sequence data safely stored on your local machine. It even features a built-in local chat assistant that can write scripts and automate your cloning design work. It is the ultimate lightweight playground for anyone curious about bioengineering.

🆔 @hackernewsgithubprojects