📦 doriandarko/texts-to-transformer

Texts to Transformer

You can train a custom language model entirely on your own iMessage history right on your Mac, and you do not need any external servers or pre-trained weights to do it. A clever project called texts-to-transformer lets you build a personal style model completely from scratch using Apple Silicon. It securely snapshots your local database, pseudonymizes sensitive identifiers, and trains a tiny decoder-only transformer that perfectly captures your slang, phrasing, and texting rhythm. It is a brilliant way to see how language models actually learn to mimic a specific voice, all while keeping your private conversations safely on your device.

🆔 @hackernewsgithubprojects

📦 wolfiesch/sf-parking-heatmap

SF Parking Heatmap

San Francisco's parking meters processed over two hundred million transactions, and this interactive map turns that massive dataset into a weekly occupancy predictor for every single metered block in the city. By crunching ninety days of transaction data into a clean, hour-by-hour profile, sf-parking-heatmap lets you scrub through any day and time to see where finding a spot is actually possible. It visualizes parking pressure as 3D columns and heatmaps, overlays bike share demand, and even calculates drive time reachability using local routing. It is the ultimate tool for understanding the rhythmic heartbeat of urban parking. Check it out to see how big data can make city driving just a...

🆔 @hackernewsgithubprojects

📦 eatmoreduck/boss-zhipin-scraper

boss-zhipin-scraper

boss-zhipin-scraper is the job-hunting tool that successfully bypasses annoying font-scrambling traps to deliver clean, readable salary data. By connecting directly to an active Chrome browser via the developer console protocol, this script leverages your existing login session to query the platform's search API. It pulls back pure text job details, saves them into neat JSON or CSV files, and even generates custom resume-tuning prompts. Grab the exact numbers, analyze required skills, and stop squinting at uncopyable job descriptions. Clone it today to quickly turn chaotic online listings into organized spreadsheets and find your next opportunity.

🆔 @hackernewsgithubprojects

📦 internrobotics/internnav

internnav

internnav is the open-source toolbox that finally bridges the gap between simulated robotic training and real-world navigation. If you are building smart robots, you know how painful it is to make them move and understand commands at the same time. This project gives you a unified playground built on PyTorch to train robots using vision and natural language instructions. It supports major simulation tools like Habitat and Isaac Sim, meaning you can easily test your models before deploying them onto physical hardware. Check out internnav to build smarter robots that actually understand where they are going.

📰 https://news.ycombinator.com/item?id=48855572

🆔 @hackernewsgithubprojects

📦 wenyi401/ikik-api

ikik-api

Pool all your different AI accounts into a single, unified gateway. This clever project lets you combine subscription accounts and private API keys from major providers like OpenAI, Claude, Gemini, and Grok into one shared, self-hosted pool. Instead of juggling a dozen separate dashboards, API keys, and subscription limits for different tools, you get a single endpoint that handles smart routing, usage tracking, and member billing automatically. It is the ultimate way to set up a private AI carpool for your team or friends so everyone can access premium models without wasting money on duplicate individual subscriptions. Check it out and simplify your AI access today.

🆔 @hackernewsgithubprojects

📦 sb-ocr/cmdeck

CM Deck

Build your own pocket-sized folding computer using a custom clamshell design. CM Deck is a fully open-source cyberdeck designed around the Raspberry Pi Compute Module 5. It features a custom-engineered carrier board paired with a low-profile mechanical keyboard running specialized open-source firmware. Instead of relying on bulky, off-the-shelf parts, this project gives you the actual hardware layouts and files needed to assemble a sleek, integrated portable machine from scratch. It is a fantastic template for anyone wanting to break away from standard laptops and craft a highly personalized, tactile computer that fits right in a bag.

📰 https://news.ycombinator.com/item?id=48855708

🆔 @hackernewsgithubprojects

📦 kimsh-1/gongnyang-prompt-kit

gongnyang-prompt-kit

gongnyang-prompt-kit is the clever Claude Code skill that turns your vaguest, one-sentence ideas into production-ready image generation prompts. If you have ever asked an AI generator for a poster and got back hot garbage, you know the pain of writing long, painful camera angles, hex codes, and layouts. This tool automatically compiles your lazy requests into structured prompts complete with scenes, lighting, and aspect ratios. It even runs a built-in validator to check the output before you generate. It is like having a professional designer translate your messy thoughts into beautiful, predictable art on the very first try.

🆔 @hackernewsgithubprojects

📦 bino5150/lumina

Lumina: The 100% Local AI Agent Built to Fit on Consumer Hardware

Most local AI assistants are bloated, memory-wiping wrappers that will choke your hardware, but Lumina is a complete, privacy-first agentic operating system built from scratch to run entirely on your own machine. Designed on a modest four-gigabyte graphics card, this clever tool bypasses corporate cloud servers completely. It features a unique three-layer memory system that actually remembers your projects and decays naturally over time, and it even reviews its own day to dream up new memories while idle. With native desktop controls, voice cloning, and sixty built-in tools, it is the ultimate loyal companion for your desktop.

🆔 @hackernewsgithubprojects

📦 lsxi7/wat3r

Wat3R

The underwater mapping tool that maps deep-sea geometry without needing any manual data labels, wat3r is changing how we reconstruct the ocean floor. Usually, teaching an AI to map things in 3D requires thousands of hand-labeled images, which is basically impossible to get underwater because of murky, scattering light. This project solves that by adapting models trained on dry land and training them on raw, unlabeled ocean videos. It uses a clever cross-view trick to piece together depth maps and camera positions, cutting through the haze. You can plug in normal underwater photos and instantly get 3D point clouds.

🆔 @hackernewsgithubprojects

📦 pengjie-w/deltav

deltav

An artificial intelligence model can now reason through complex visual tasks by only tracking what changes from one step to the next, rather than regenerating a complete image every single time. This is the core breakthrough of deltav, a unified large multimodal model designed for step-by-step visual reasoning. Traditional systems waste massive amounts of computation by repeatedly processing static, unchanged backgrounds during a logical sequence. Instead, this project uses a clever routing system that measures visual changes over time and dynamically updates only the reasoning-critical modifications. It is a highly efficient way to help computers solve visual puzzles, plan spatial layouts, and handle multi-step logic without drowning in redundant data.

🆔 @hackernewsgithubprojects

📦 fduflyer/droneaware-node-releases

droneaware-node-releases

droneaware-node-releases is the community-driven software that turns a basic Raspberry Pi into a real-time drone tracking station. By connecting a cheap USB Bluetooth and Wi-Fi adapter, your DIY node scans the airwaves for mandatory drone Remote ID broadcasts, acting like a local radar for unmanned aircraft. It decodes the drone’s identity, speed, and flight path, forwarding the data to a live global map. You can monitor everything from your home network, receive instant email alerts when a drone flies nearby, and even watch detailed flight replays. It is a brilliant way to bring airspace transparency right to your neighborhood.

🆔 @hackernewsgithubprojects

📦 shengshu-ai/vidu-s1

Vidu S1

Vidu-s1 is the AI model that finally lets you talk to digital characters and watch them react in real time. Instead of waiting minutes for an AI video to render, this project lets you verbally instruct a custom avatar, an anime character, or even a pet, and watch them perform actions instantly. It generates standard video at a blazing fast forty-two frames per second, running smoothly on normal consumer graphics cards. It basically turns video generation into a live, interactive conversation. If you want to see how fast video AI is moving, you should definitely check this out.

🆔 @hackernewsgithubprojects

📦 velornlabs/velorn

Velorn: The AI Video Editor Powered by ComfyUI

Turn your ComfyUI generation chaos into a polished, structured masterpiece with a timeline editor built directly for AI creators. Velorn brings your local and cloud workflows into a single workspace, letting you assemble music videos from lyrics, generate UGC-style ads, and trim clips with multi-track control. Instead of jumping between endless tabs and file folders, you can run custom workflows, generate assets, apply transitions, and style captions all in one place. It even includes a local agent server that can inspect your timeline and troubleshoot missing models. It is the ultimate production layer to finally finish your AI video edits.

🆔 @hackernewsgithubprojects

📦 mudler/depth-anything.cpp

Depth Anything CPP

This lightweight tool can actually compute incredibly detailed 3D depth maps on your computer's CPU faster than PyTorch, all without needing Python or a massive deep learning framework. It is called depth-anything.cpp, and it is a pure C++ port of ByteDance's depth-sensing model that runs entirely inside a single, self-contained file. By rewriting the code from scratch and caching the heavy geometric calculations, this engine uses half the memory of PyTorch and loads almost seven times faster. You can extract 3D point clouds, camera angles, and depth maps from a simple photo in a fraction of a second.

🆔 @hackernewsgithubprojects

📦 zju-omniai/vla-corrector

VLA Corrector

vla-corrector is the robot control framework that finally stops robotic arms from repeating their own mistakes. Traditional robot AI models plan out a long list of actions and then blindly execute them, completely ignoring if a gripper slips or an object slides away mid-task. Instead of throwing away your massive AI models, this project plugs in a lightweight, forty-million-parameter monitor that watches the camera feed in real time. The moment it detects a mismatch between what the robot expects to happen and what is actually happening, it instantly stops the stale action queue and recalculates a recovery plan. Grab the code on GitHub to give your robots actual common sense.

🆔 @hackernewsgithubprojects

📦 labrai/skillcenter

skillcenter

skillcenter is the clever open-source indexer that packages over one hundred thousand ready-to-use AI agent skills into a single local command. Instead of manually writing custom prompts or hunting down setup instructions for your AI coding assistants, this tool lets you search and install tailored skill files directly into your project in seconds. It crawls thousands of repositories, research papers, and developer sources, packing them into an offline SQLite database you can search instantly. You just run a quick command to find what you need, install it to your project, and your agent immediately picks up the new capability. Check out skillcenter on GitHub to supercharge your local developer agents...

🆔 @hackernewsgithubprojects

📦 sarthaxxxxx/awesome-continual-test-time-adaptation

Awesome Continual Test Time Adaptation

awesome-continual-test-time-adaptation is the curated directory that finally maps out how AI models can adapt to the real world on the fly. Imagine deploying a smart camera, but then it rains, the lighting shifts, or the camera lens gets dusty, causing the AI to completely freeze up. This repository gathers all the research, benchmarks, and strategies for updating AI models in real-time as they run, without erasing their original training or needing expensive servers. It is like giving an AI model a pair of glasses that constantly adjusts its focus so it never gets confused by a little weather change. Check it out to see how the smartest models stay sharp.

🆔 @hackernewsgithubprojects

📦 kuskhan/jetendard

Jetendard

Jetendard perfectly solves the awkward spacing issues that usually happen when combining English coding fonts with Korean characters. By combining JetBrainsMono Nerd Font with Pretendard Korean glyphs, this clever python-based builder scales the Korean characters up by exactly fifteen percent. This slight adjustment removes the unnecessary gaps around Korean words, aligning them to exactly twice the width of the English characters. The result is an incredibly crisp, visually stable, and highly legible monospace font that gives developers writing in Korean a beautiful and perfectly proportioned coding environment in their terminals and editors.

🆔 @hackernewsgithubprojects

📦 sbroenne/mcp-server-excel

mcp-server-excel

Connect your favorite AI assistant directly to Microsoft Excel to automate spreadsheet workflows using everyday language. This clever project bridges artificial intelligence and your local Excel application using the Model Context Protocol. Unlike tools that only parse static files, it actually drives your installed Excel app in real time, letting an AI build formulas, refresh Power Query, run VBA macros, and update pivot tables on the fly. You can watch your spreadsheet update instantly as the AI works. It is the perfect playground for automating tedious data tasks without writing complex scripts. Give it a spin and let your AI do the heavy lifting.

🆔 @hackernewsgithubprojects

📦 davian-robotics/3d_hamster

3D HAMSTER

3D HAMSTER is the depth-aware robot planner that finally keeps your machine's trajectory grounded in physical reality. While typical AI planners only figure out movements on a flat two-dimensional screen, this system calculates real physical coordinates in space directly from a single camera feed and a simple depth map. By planning in actual metric measurements, it makes sure the robot's arm never gets confused by flat pixels or makes clumsy physical mistakes. It acts as a bridge between high-level language commands and actual hardware controls. Grab the open-source code and download the ready-to-use model weights to start teaching your robots to move with real physical awareness today.

🆔 @hackernewsgithubprojects