📦 svg-project/quant-videogen

Quant VideoGen

Compress the massive memory footprint of generating long AI videos by up to eighty-five percent without changing your model weights or sacrificing quality. This clever project tackles the heavy memory cost of long-horizon attention data during video generation by quantizing the temporary cache down to just two bits during inference. This simple step slashes the memory needed for this cache by about seven times, turning a massive hardware bottleneck into an afterthought. It works seamlessly across several popular open-source video models to let you run longer creations on much more modest hardware. Give it a try to experience incredibly efficient long-context video generation.

🆔 @hackernewsgithubprojects

📦 shannhk/llm-wikid

llm-wikid

Build a self-growing second brain inside Obsidian by dumping raw articles, tweets, and PDFs into a folder and letting an AI agent compile them. Instead of slow vector database lookups, llm-wikid uses an agent to auto-sort your clippings, fetch full web content, download diagrams, and compile structured wiki pages with active cross-references. It even runs bias checks to highlight conflicting viewpoints and data gaps. Every question you ask gets answered with citations and filed back in, making your knowledge vault smarter the more you use it. Hook up your favorite agent, clone the repository, and start building a highly organized personal encyclopedia today.

🆔 @hackernewsgithubprojects

📦 flavioadamo/openfit

OpenFit

The Fitbit desktop dashboard that finally puts your personal health data back in your own hands. OpenFit is a private, desktop-first app that pulls your activity, sleep, and heart metrics directly from Google Health into a beautiful, local interface. Instead of staring at empty charts, its clever layout completely adapts to only show the metrics your device actually recorded. It even includes a secure, local AI assistant to chat about your health trends without ever sharing your credentials. It is the perfect, secure way to visualize your fitness journey on a bigger screen over a morning coffee.

🆔 @hackernewsgithubprojects

📦 lzh-phd/topic-feasibility-screener

Topic Feasibility Screener

Turn a rough research idea and a raw dataset into a complete, visual academic dashboard. Topic-feasibility-screener takes your CSV or Stata file, brainstorms dozens of concrete research questions, and automatically searches real academic databases to find genuine literature gaps. Instead of just chasing tiny p-values, it runs rigorous statistical checks like panel fixed effects and parallel trend tests, blending empirical strength with actual theoretical backing to score and rank your ideas. It hands you a beautiful HTML report packed with interactive feasibility charts and structured research designs so you can instantly see which paper topic is actually worth writing.

🆔 @hackernewsgithubprojects

📦 keyvaluesoftwaresystems/agent-opfor

Agent Opfor

Agent-opfor is the open-source adversary emulation library that tests your AI agents and servers by actively attacking them before real bad actors can. It simulates realistic, multi-turn conversations and sneaky maneuvers, mapping directly to security standards to expose weaknesses in tools, prompts, and memory. The coolest part is how it lets you see behind the curtain. By plugging into your monitoring stack, the evaluator doesn't just watch what your agent says, it inspects every hidden tool call and database check to catch silent data leaks. It is an incredibly clever way to secure your AI.

🆔 @hackernewsgithubprojects

📦 xiaohuailabs/xiaohu-ip-studio

Xiaohu IP Studio

Illustrate your deep Chinese articles with a consistent cartoon character acting out your complex ideas instead of relying on boring bullet points and generic clip art. This project acts as an automated studio that reads your text, identifies the perfect paragraphs for visual metaphors, and designs fresh illustrations using a library of thirty-one built-in hand-drawn characters and meme mascots. It translates dry mechanics into emotional hooks, explanatory diagrams, or even four-panel comics, running self-checks to ensure the text and characters look perfect. You can install it as a smart agent skill to build a recognizable identity for your content instantly.

🆔 @hackernewsgithubprojects

📦 lab421/forel

Forel

Organize your messy macOS folders automatically in the background without sending a single byte of your data to the cloud. This clever tool lets you set up smart rules that monitor folders in real time, matching your files by size, date, or extension, and then instantly moves, renames, or tags them. The coolest part is that it can actually read text inside scanned PDFs and images using local, on-device machine learning, so you can sort receipts or invoices completely offline. It runs quietly in your menu bar, keeping your desktop clean and your private files totally private.

🆔 @hackernewsgithubprojects

📦 k-dense-ai/mimeo

Mimeo

Clone an expert's way of thinking directly into your AI coding assistant. This clever tool automatically reads the internet on your behalf to distill the exact mental models, principles, and habits of industry leaders like Andrej Karpathy or Andrew Ng. It crawls essays, lectures, and interviews, cleans and verifies the sources, and then packages everything into a clean markdown file that your coding agent can read. By feeding these customized guides to your AI, you immediately install the expert's default stance, problem-solving frameworks, and design philosophies into its everyday behavior. It is a brilliant way to level up your AI assistant.

🆔 @hackernewsgithubprojects

📦 pepetii/danmuai

DanmuAI

DanmuAI is the Windows desktop companion that lets streamers run their live chat overlay entirely through local vision models. Instead of reading boring raw text, this program takes periodic screenshots of your selected display, passes them to your favorite AI provider, and instantly rolls out context-aware scrolling chat messages onto a beautiful transparent layer. Your stream key and personal data stay completely private because the app processes screenshots directly in memory without writing files to your drive. It is a brilliant, privacy-first way to make any live broadcast feel alive, responsive, and incredibly engaging.

🆔 @hackernewsgithubprojects

📦 chainstacklabs/polymarket-alpha-bot

Polymarket-Alpha-Bot

polymarket-alpha-bot is the smart prediction market assistant that identifies profitable hedge portfolios across correlated events. Instead of guessing how different prediction markets impact one another, this tool uses a clever combination of logical rules and large language models to map out real-time relationships between active topics. It automatically flags overlapping positions where you can hedge your bets, tracks live prices, and lets you execute covered pair trades directly from a clean web dashboard without dealing with manual calculations. It is a brilliant way to spot hidden pricing gaps and manage your exposure like a professional trader without needing a finance degree.

🆔 @hackernewsgithubprojects

📦 bstnxbt/dflash-mlx

DFlash MLX

Speed up your local large language models on Apple Silicon by up to four times using dflash-mlx. This clever tool brings lossless speculative decoding to your Mac, which means it uses a tiny, hyper-fast draft model to predict sixteen tokens at once, and then lets your main model verify them in a single pass. If any predicted tokens are wrong, it instantly rolls them back and keeps going. Because every single token is verified before it is committed, your output stays perfectly accurate. It integrates beautifully with tools you already use, giving you massive speedups without losing quality.

🆔 @hackernewsgithubprojects

📦 us/crw

crw

Turn any messy website into clean, structured data for your AI models in milliseconds. The open-source tool crw is a lightning-fast web scraper and crawler built specifically for AI agents, and it runs as a single, lightweight binary that uses just fifty megabytes of memory. Instead of wrestling with heavy, browser-heavy setups that hog your server resources, you can self-host this engine on a cheap virtual private server to scrape, crawl, and track website changes using a simple API. It is the perfect way to feed fresh, clean markdown directly into your LLMs without the headache. Give it a spin and watch your AI agents get way smarter.

📰 https://news.ycombinator.com/item?id=48819736

🆔 @hackernewsgithubprojects

📦 jmstar85/oh-my-githubcopilot

oh-my-githubcopilot

This project changes how you use AI in VS Code by bringing a coordinated multi-agent system directly into GitHub Copilot. Instead of relying on a single general assistant, it deploys twenty-eight specialized agents and thirty-seven reusable skills managed through a local model context protocol server. This setup separates planning, implementation, and code review, using built-in safety guardrails to block destructive commands while enforcing verification and automated testing. It is a smart way to turn a basic chat prompt into a structured, reliable development pipeline. Check out the repository to level up your programming workflow today.

🆔 @hackernewsgithubprojects

📦 arahim3/mlx-dspark

mlx-dspark

Your Mac can now run DeepSeek's advanced speculative decoding technology to generate local text up to sixty percent faster without losing a single drop of quality. The mlx-dspark project brings both DeepSeek's DSpark and z-lab's DFlash architectures natively to Apple Silicon using Apple's MLX framework. Instead of waiting for a large language model to think up every word sequentially, this tool uses a tiny, hyper-fast draft model to guess ahead, which the main model verifies in a single parallel pass. It even serves an OpenAI-compatible API with continuous batching and prefix caching, making your local offline assistant feel incredibly snappy.

🆔 @hackernewsgithubprojects

📦 howdymary/autopredict

autopredict

Build and backtest smart trading agents for prediction markets like Polymarket, and let them automatically improve their own strategies over time. This clever little Python framework scans live market data and runs historical backtests, but its coolest feature is a self-improvement loop that actually tests and refines your trading rules. It rolls out strategy tweaks, grades them against real historical outcomes, and only saves the new version if it legitimately beats the old one. There is no fake data or magic default formulas here, just an honest tool to help your code learn how to trade better on its own.

🆔 @hackernewsgithubprojects

📦 ruc-datalab/deepanalyze

DeepAnalyze

An open-source artificial intelligence system can now act as your personal data scientist, managing complex data projects completely on its own. This new project, called deepanalyze, automates the entire data science pipeline, from preparing raw files to training predictive models and generating polished, professional reports. Instead of forcing you to clean data and write repetitive code, it autonomously explores structured spreadsheet tables, nested files, and plain text documents to find insights. It runs python code in sandboxed environments to verify its work, meaning you get fully validated analysis. It is a powerful way for anyone to turn messy folders of raw files into clear, visual research reports instantly.

🆔 @hackernewsgithubprojects

📦 larashero3-dotcom/writing-dna-skill

writing-dna-skill

Artificial intelligence can copy any writing style consistently by breaking down texts into structured, reusable blueprints instead of just guessing the tone from a few examples. The writing-dna-skill repository provides a comprehensive system that analyzes at least twenty sample articles across six distinct layers, separating style from mere text. It extracts a writer's specific vocabulary, sentence rhythms, structural templates, core thinking frameworks, and even visual formatting guidelines. This structured blueprint is then saved as a single master file that an AI agent can reread and reference before drafting new pieces, ensuring the output perfectly mirrors the original writer's voice, depth, and presentation.

🆔 @hackernewsgithubprojects

📦 ostris/comfyui-krea2-ostris-edit

comfyui-krea2-ostris-edit

The ComfyUI custom node extension that finally brings reference-based image editing to the Krea 2 model. While Krea 2 is normally a text-to-image generator, this tool patches the model so it can actually consume reference images alongside your prompts. It encodes your images through a vision-based text encoder and feeds them straight to the model at the very beginning of the generation. This means you can train a custom edit style, feed in a reference photo, and transform it using specialized edit adapters. It is the perfect bridge for anyone wanting to bend Krea 2 into a powerful, controllable editor.

🆔 @hackernewsgithubprojects

📦 v4bel/januscape

Januscape

Crash any multi-tenant x86 cloud server directly from inside your virtual machine using a newly discovered exploit. The project, called januscape, showcases a critical guest-to-host escape vulnerability that lay dormant in the Linux kernel for sixteen years. By abusing a logic flaw in how KVM handles memory virtualization, a guest virtual machine can trick the host kernel into reusing memory structures with the wrong architectural roles, corrupting the host's memory state. This allows an attacker renting a single cloud instance to trigger an immediate host kernel panic, instantly taking down every other virtual machine on that physical machine.

📰 https://news.ycombinator.com/item?id=48807908

🆔 @hackernewsgithubprojects

📦 robbyant/lingbot-vision

LingBot-Vision

Lingbot-vision is the vision artificial intelligence model that maps boundaries and shapes directly from raw images without needing human-labeled training data. By focusing on the edges where objects meet, this system naturally grasps the physical layout of a room, including tricky reflective glass and deep spaces. It acts like a digital eye that groups pixels into coherent shapes, helping computer vision programs accurately measure depth, trace object contours, and track moving things in video. It is a incredibly smart way to give robots and cameras a spatial common sense, making visual AI dramatically better at understanding the physical world.

🆔 @hackernewsgithubprojects