📦 dekart-xyz/geosql

Geo SQL

Your AI assistant can now write perfect spatial queries and catch its own coding errors by looking at a map. An open source tool called geosql acts as a bridge that installs directly into local AI tools like Claude, Codex, and GitHub Copilot. It explores your actual database schemas, drafts precise geographical queries, and enforces smart cost guards so you never accidentally run a massive, expensive query. But here is the brilliant part: it spins up a local map, looks at the visual output, and uses that visual feedback to self-correct boundary mistakes that text alone would miss. It is basically giving your AI a pair of eyes for geographic...

📰 https://news.ycombinator.com/item?id=48829242

🆔 @hackernewsgithubprojects

📦 robodojo-benchmark/robodojo

RoboDojo

Evaluate generalist robots on complex, hard manipulation tasks using a standardized testing ground. RoboDojo provides a unified framework to test robot skills across fifty-four challenging simulation and real-world tasks, targeting crucial capabilities like generalization, memory, and precision. By separating the physics simulation from the control policy, you can plug in over forty different AI models and evaluate them under identical conditions. It brings together physically grounded objects, multi-task parallel runs, and a consistent leaderboard. It is the ultimate playground for seeing how smart a robot controller really is before deploying it to the physical world.

🆔 @hackernewsgithubprojects

📦 rightnow-ai/auto

Auto

Compile recorded AI agent behaviors into fast, secure, and incredibly cheap WebAssembly binaries. The auto repository treats expensive frontier models as interpreters, records how they handle tasks, and extracts the underlying logic. It figures out which parts of an agent's reasoning are actually symbolic patterns and compiles them into sandboxed binaries that run on small, specialized models. If the system encounters a brand-new scenario, a guard automatically hands the task back to the frontier model, records the new solution, and recompiles it. This means you never pay to solve the same complex problem twice, cutting operational costs while keeping execution blazing fast.

📰 https://news.ycombinator.com/item?id=48822488

🆔 @hackernewsgithubprojects

📦 powertony102/recal3r

recal3r

recal3r is the streaming three-D reconstruction tool that finally keeps long video-to-three-D models from falling apart over time. If you have ever tried turning a video stream into a three-D map, you know that mistakes build up quickly, creating messy, warped reconstructions. This project solves that by analyzing how reliable each part of the video's data is during the reconstruction process. It automatically scales down the training rate for uncertain areas so noisy data does not ruin the model's memory. The result is a much cleaner, highly accurate three-D reconstruction over long video paths, giving you precise camera tracking and depth mapping.

🆔 @hackernewsgithubprojects

📦 showlab/padcaptioner

PadCaptioner

AI models are notoriously slow at writing detailed captions for long videos because they generate descriptions word by word, in a single file line. PadCaptioner changes the game by proving a smaller model can write detailed video descriptions faster and better than models twice its size. Instead of waiting for one sentence to finish before starting the next, it figures out which events in a video don't depend on each other and generates those captions in parallel. It uses a clever planning step to group distinct moments together, giving you lightning-fast, highly accurate video summaries. Keep an eye on this project, as the code is dropping soon.

🆔 @hackernewsgithubprojects

📦 kklimuk/docx-cli

DOCX CLI: The Ultimate Tool for AI Agents Editing Word Documents

Handing a Word document to an AI agent usually ends in absolute disaster, with the model bloating your token count or spitting out corrupted XML files that Word refuses to open. That is where docx-cli comes in, letting AI agents seamlessly read, edit, and redline Word documents from the command line while keeping the styling completely intact. By converting the document into annotated Markdown with stable locators and applying edits directly to the underlying XML, the agent never has to struggle with formatting. This clever tool is a must-have for building robust, document-editing AI workflows.

📰 https://news.ycombinator.com/item?id=48821500

🆔 @hackernewsgithubprojects

📦 oceanflowlab/omnivtg

OmniVTG

Pinpoint exact moments in long, unedited videos just by typing a simple text description. OmniVTG introduces a massive dataset featuring over two thousand hours of diverse video content alongside a training system that teaches artificial intelligence how to think through video searches. Instead of just guessing, the AI uses a self-correction chain of thought to double-check its work and find precise start and end times down to the second. It solves the frustrating challenge of hunting through raw footage, making open-world video search incredibly accurate. Check out the project to see how it can find any action in seconds.

🆔 @hackernewsgithubprojects

📦 alibaba-damo-academy/rynnworld-4d

Rynn World 4D

Teach your robots how to interact with the physical world by predicting geometry and movement simultaneously. rynnworld-4d does this by generating synchronized video, depth, and optical flow inside a unified four-dimensional simulation. Instead of just guessing what flat pixels will look like next, this system tracks how objects actually move and change shape in 3D space. It uses a clever three-branch neural network to keep the appearance, depth, and motion perfectly aligned, which feeds directly into an action controller to guide dual-arm robot hands in real time. It is a brilliant way to give machines a better sense of physical intuition.

🆔 @hackernewsgithubprojects

📦 lqs/retrotick

RetroTick

You can actually run retro Windows and DOS executables directly in your web browser without running a heavy virtual machine. A brilliant project called retrotick accomplishes this by emulating an x86 processor and rebuilding classic operating system interfaces entirely in TypeScript. Instead of loading a massive virtual disk, you just drag and drop an old application right onto the page. It translates old graphics and audio on the fly, meaning you can play original games like Minesweeper or launch classic 3D screensavers in a web tab. It is an incredibly clever way to preserve vintage software using modern web standards.

🆔 @hackernewsgithubprojects

📦 sakuraaa0/rabitqcache

Rabitqcache

Running long-context AI models usually slows to a crawl because storing the conversation history eats up massive amounts of GPU memory. Rabitqcache fixes this by compressing that history down to tiny binary representations using randomized rotation and high-speed math. Instead of guessing which past tokens to look at with expensive calculations, this system uses an incredibly fast math trick to dynamically grab only the exact history it needs on the fly. It even hides the compression process in the background so your hardware never has to wait around. If you want to run massive prompt histories on normal hardware without your generation speeds falling off a cliff, you should definitely...

🆔 @hackernewsgithubprojects

📦 ruvnet/ruvector

RuVector

RuVector is the self-learning vector database that actually gets smarter the more you use it. While standard databases just search the same way forever, this clever tool watches your queries to automatically improve search results and tune itself on the fly. It runs AI models locally on your own hardware without any expensive cloud APIs, and it can drop right into PostgreSQL or run inside a web browser from a single file. It is basically a brain for your data that fits anywhere. You get better, faster search that adapts to your workflow without the headache of managing complex cloud clusters.

🆔 @hackernewsgithubprojects

📦 johnviljoen/jaxipm

jaxipm

An open-source tool called jaxipm can solve hundreds of complex mathematical optimization problems at the exact same time on a single graphics card. Typically, hardware-accelerated math solvers run tasks one after another, which leaves valuable computing power sitting idle. This project solves that waste by bundling those separate problems into batches, fully saturating the processor's capacity. By utilizing specialized hardware libraries to handle these groups simultaneously, it achieves massive speedups compared to traditional methods. It shows that we do not need massive supercomputers for heavy optimization tasks when we can coordinate our existing hardware more efficiently.

🆔 @hackernewsgithubprojects

📦 timjaspers0801/atlas

Atlas

Teach computers to recognize real-time human anatomy during surgery by setting up a massive medical imaging pipeline. This project provides the downloading and processing tools for a massive dataset of over one hundred and twenty thousand annotated video frames spanning fourteen different surgical procedures. Instead of forcing you to store massive video files, it lets you download public surgical videos and automatically extract only the precise frames containing key organs and tools, perfectly aligned with expert-labeled anatomical masks. It is a brilliant way to train medical artificial intelligence without drowning in raw video storage, giving you clean, curated data for smarter surgical assistants.

🆔 @hackernewsgithubprojects

📦 brain0-ai/brain0

Brain0

Brain0 is the black box for AI-written code that connects every single commit back to the exact agent prompts that wrote them. While git shows you what changed, this tool watches your workspace offline and links actual code modifications to the original prompts and agent logs. It even flags drift when an AI claims it changed one thing but secretly touched another, and audits whether your coding agents read sensitive files before talking to remote servers. It is like an automated flight recorder for your codebase that gives you real, evidence-driven risk scoring without slowing down your existing development loop.

🆔 @hackernewsgithubprojects

📦 ninjahawk/subtext

Subtext

Subtext is the real-time visualizer that finally exposes the hidden, unspoken thoughts of an artificial intelligence as it reads and types. Normally, we only see the final words a language model outputs, but this project taps directly into the model's inner mind using a diagnostic lens. By running a local model and reading its neural activity at multiple depths, Subtext captures concepts before they are spoken. For example, if you ask the model to check an incorrect math equation, the word incorrect instantly lights up in its internal workspace while it is still reading your prompt, several steps before it actually starts typing a reply. It lets you watch the...

📰 https://news.ycombinator.com/item?id=48811892

🆔 @hackernewsgithubprojects

📦 vbiroshak/ai-project-architect

AI Project Architect

ai-project-architect is the workspace framework that finally solves AI memory loss by giving your local assistant its own persistent filesystem diary. Instead of letting your AI helper forget everything the moment you close a chat, this system sets up a structured directory on your local drive that the AI reads, writes, and updates directly. It tracks its own tasks, manages an inbox of new files, and even uses a clever file-timestamp trick to know exactly how much time has passed since your last session. You get a portable, local knowledge base that naturally picks up right where you left off.

🆔 @hackernewsgithubprojects

📦 sunflower-of-parchman/codex-hygiene

Codex Hygiene

Codex-hygiene is the local agent skill that finally makes it easy to monitor and clean up your Codex Desktop environment without losing your active workflows. If you have ever noticed your AI assistant lagging or burning through tokens, it is usually because of hidden junk like bloated tool lists, forgotten background plugins, or massive chat histories reloading behind the scenes. This clever utility uses safe, read-only database queries to pinpoint exactly what is eating up your context window. It measures your active memory footprint and suggests reversible settings tweaks, giving you a clean slate while keeping your core project goals completely intact.

🆔 @hackernewsgithubprojects

📦 evolvinglmms-lab/skillopt-lite

skillopt-lite

AI coding agents can now automatically train themselves to get better at complex tasks without you writing a single line of training code. An open-source project called skillopt-lite lets you point your favorite AI coding assistant at your workspace and trigger an optimization loop with a single chat command. The tool guides your agent to run evaluations, analyze its own failures, update its natural-language instructions, and roll back any changes that decrease performance. It handles the entire trial-and-error cycle of self-improvement completely inside your editor chat. Just run the loop, sit back, and let your agent polish its own skills.

🆔 @hackernewsgithubprojects

📦 iamcheyan/fudoki

Fudoki

Fudoki automatically breaks down complex Japanese text to instantly show you parts of speech, readings, and translations directly in your browser. If you are learning Japanese, trying to read native material can feel like running into a wall of kanji with no idea where words begin or end. Fudoki solves this by instantly segmenting sentences, color-coding nouns, verbs, and adjectives, and overlaying phonetic readings and English definitions. It even includes a built-in markdown editor and high-quality speech synthesis so you can listen to any word or sentence read aloud at your own pace. It is like having a patient, sharp-eyed language tutor built right into your browser.

🆔 @hackernewsgithubprojects

📦 scaleapi/scale-agentex

scale-agentex

Build and scale intelligent AI systems that can run complex, autonomous tasks instead of just waiting for simple back-and-forth chat responses. While most AI applications are stuck handling quick, basic questions, scale-agentex gives you a complete framework to run advanced, long-running agent workflows. It lets you spin up a local development server, try out pre-made templates, and test everything instantly in a built-in web dashboard. It is a fantastic way to experiment with multi-agent setups that actually work in the real world. Get your first system running locally and watch your AI take on real work.

📰 https://news.ycombinator.com/item?id=48831782

🆔 @hackernewsgithubprojects