📦 escoffier-labs/brigade

Brigade

Artificial intelligence agents can execute complex loops and generate code, but keeping track of what they actually accomplished is notoriously difficult. Brigade solves this by acting as a lightweight, local control plane that creates verifiable file receipts for every single action your agents take. Running completely on your own machine without background daemons or platform lock-in, it lets different AI tools and frameworks easily share the Model Context Protocol, custom tools, and memory. Because it records everything with tamper-resistant receipts, you can easily prove what your agents did and use real program exit codes to systematically improve their performance.

🆔 @hackernewsgithubprojects

📦 nvidia-nemo/emerging-optimizers

Emerging Optimizers

Nvidia’s emerging-optimizers repository implements advanced training algorithms that actually speed up how fast massive AI models learn. Traditional training rely on simple adjustments, but this project packages sophisticated alternatives like Muon and Soap, which use advanced matrix operations and geometry to find the best settings much faster. In fact, these exact techniques were used to train the massive Kimi K2 language model. By integrating these tools directly into standard graphics processing unit training systems, developers can achieve quicker convergence and better final model quality. It is a powerful toolbox for anyone looking to push the boundaries of neural network efficiency.

🆔 @hackernewsgithubprojects

📦 maxwell2732/pai-econ-claude

pai econ claude

Empirical economists often struggle to bridge the gap between real-world data and formal mathematical modeling. The open-source project pai-econ-claude solves this by acting as a highly structured theoretical scaffolding for empirical researchers. Instead of building models from scratch, it matches real-world economic puzzles with classic theoretical frameworks like search models, moral hazard, or rational inattention. The tool walks users through rigorous checks, auditing assumptions, outlining proof sketches, and organizing findings into draft manuscripts. It is a brilliant way to make sure your empirical insights are grounded in solid, peer-reviewed economic theory before you write your paper.

🆔 @hackernewsgithubprojects

📦 shanghai-academy-of-ai-for-science/mkb

mkb

Imagine a single AI model that can design RNA sequences, forecast global weather, and segment medical images with expert-level precision. That is exactly what mkb does. Built on an eight-billion parameter language backbone, this system pairs specialized scientific encoders and decoders with a shared core. Instead of swapping out different models for different tasks, it routes biology, chemistry, and physics data through a single natural-language interface. It actually outperforms massive one-trillion parameter models on biological tasks and rivals physics-based forecasting systems. It is a brilliant, lightweight way to solve diverse scientific problems on a single graphics card.

🆔 @hackernewsgithubprojects

📦 ibm-client-engineering/output-drift-financial-llms

Output Drift Financial LLMs

Stop relying on massive AI models for regulated tasks without checking their consistency first. This repository lets you benchmark and audit how much large language model outputs drift when running financial operations like text to SQL or compliance triage. Interestingly, testing reveals that smaller seven to twenty billion parameter models achieve up to one hundred percent consistency, while massive frontier models over one hundred billion parameters show only twelve to fifty percent consistency. You can run repeated trials on identical prompts to measure action, signature, and decision determinism to ensure your automated agents are completely reproducible before they ever touch live financial data.

🆔 @hackernewsgithubprojects

📦 joeseesun/qiaomu-cut-skill

Qiaomu Cut Skill

You can now transform a single spoken prompt into a complete, ready-to-render video project using Qiaomu Cut Skill. This agent-native video director builds structure out of chaos by taking your simple prompt and generating a fully realized video timeline complete with shot lists, background music, and professional three-layer bilingual subtitles. Instead of promising magic, it coordinates actual engineering tasks, searching free stock sites, generating speech overlays, applying precise cinematic transitions, and organizing everything into structured files. It even includes a three-tiered rendering pipeline so you can quickly preview drafts before exporting your final high-quality master copy. It makes automated video editing structured and fully verifiable.

🆔 @hackernewsgithubprojects

📦 nexu-io/codex-slides

Codex Slides

Turn a prompt, a repository, or a folder of local files into a beautiful, presentation-ready slide deck without ever leaving your coding workspace. codex-slides is an open-source, image-native slide studio built for coding agents that serves as a self-hosted alternative to standard AI presentation tools. Instead of waiting for a hidden background task, you can watch the entire process live as the agent conducts research, builds an outline, and styles your project. The absolute best part is fast mode, which renders over ten slides in parallel in just four minutes. When you are done, you can edit pages on a live canvas and export your work as a PowerPoint or...

📰 https://news.ycombinator.com/item?id=49031776

🆔 @hackernewsgithubprojects

📦 saddam213/amuseai

Amuse AI Local Generator

You can now run massive artificial intelligence pipelines locally on your machine without dealing with broken Python dependencies. Amuse AI acts as a smart interface for the Tensor Stack software development kit, giving you a smooth desktop app to generate images, edit videos, and translate speech entirely on your own hardware. The coolest part is how it handles different hardware by automatically launching isolated Python environments for each specific graphics card or pipeline. This means your Nvidia and AMD setups will never clash, and you can even shrink massive models down using built-in quantization to fit them onto budget graphics cards without running out of memory.

🆔 @hackernewsgithubprojects

📦 joeyvansommeren/journey-mapper

Journey Mapper

You can now instantly turn any codebase into a beautiful, interactive customer journey map without doing a single pixel of manual design work. Journey-mapper is an automated tool that scans your routes, components, and API calls to map out exactly how users move through your application. It acts like a service designer, automatically building a single, browser-ready HTML file that combines a customer journey with a service blueprint. The tool maps out what the user does, what they see on the frontstage, and what the code triggers backstage, while marking its emotional and cognitive insights as assumptions for you to validate later. It is the easiest way to bridge the...

🆔 @hackernewsgithubprojects

📦 ailinone/collective-intelligence

Collective Intelligence

Collective Intelligence is the open source engine that lets over seventy-six thousand AI models collaborate inside a single system rather than routing requests to just one place. Instead of relying on a single model that acts as a single point of failure and training bias, this project coordinates thousands of models across dozens of strategies, like blind debates, expert panels, and consensus pools. By testing models against each other and using deterministic verifiers, it achieves an outstanding ninety-seven percent accuracy on verifiable tasks, beating major standalone frontier models. It is the ultimate way to build resilient, self-healing, and highly auditable AI applications.

🆔 @hackernewsgithubprojects

📦 maomao-2001/whispera

whispera

You can now run a fully local, real-time voice assistant directly on your Windows PC without sending a single byte of your conversations to the cloud. A clever open-source desktop app called whispera orchestrates everything right on your machine. It ties together a local language model, smart voice activity detection, and incredibly fast speech-to-text to let you have natural, spoken conversations. When you speak, it listens and can even be interrupted mid-sentence just like a real person. It generates streaming audio replies locally, and you can even plug in long-term memory so it actually remembers your past chats.

🆔 @hackernewsgithubprojects

📦 tencent/workbuddy-bench

WorkBuddy Bench

Drop your coding agents directly into a simulated office environment to see how they handle realistic corporate workloads. Rather than testing simple code snippets, workbuddy-bench runs your agent inside a local Docker sandbox to tackle complex, multi-step tasks reverse engineered from actual professional roles, like developers, product managers, and security analysts. It evaluates how effectively these agents edit real repositories, fix security vulnerabilities, modify front-end web interfaces, and manipulate messy office spreadsheet files. The system runs the agent, captures its step-by-step behavior, and scores the results automatically. It is the perfect way to pressure-test your artificial intelligence on messy, real-world tasks before letting it loose on your production code.

🆔 @hackernewsgithubprojects

📦 lexsi-labs/circuitkit

CircuitKit

CircuitKit is the mechanistic interpretability framework that lets you directly edit and export task-specialized transformer models. Most AI interpretability tools stop at showing you a colorful graph of which attention heads are active. This clever tool goes much further by finding the exact minimal circuit of neurons driving a specific behavior, letting you prune or steer that pathway, and then exporting a fully functional Hugging Face checkpoint of your newly optimized model. It makes deep model surgery accessible to anyone wanting to patch hallucinations, steer behaviors, or aggressively compress neural networks. Grab this tool and start dissecting your AI models today.

🆔 @hackernewsgithubprojects

📦 decodingai-magazine/building-a-coding-agent-from-scratch-course

building-a-coding-agent-from-scratch-course

Build a fully capable AI coding agent from the ground up without relying on complex, pre-made frameworks that hide all the magic. This repository gives you a free eight-lesson course where you construct a terminal coding assistant called decode in Python. You will go beyond the basic text generator loop to construct the actual harness that controls the model. It guides you through designing strict permission gates, running untrusted code in secure sandboxes, compressing conversation history, and even launching parallel teams of subagents to solve tasks. It is the ultimate hands-on guide to mastering how real-world AI coding tools actually operate.

🆔 @hackernewsgithubprojects

📦 role1776/netmon

Netmon

Your self-hosted network monitor can now roast your internet connection using artificial intelligence. Netmon is a brilliant little Python tool that silently watches your home network, runs hourly speed tests, and scans your local Wi-Fi to count connected devices. Every four hours, it bundles this data into a beautiful custom graph and sends it straight to your Telegram app. But the best part is the commentary. Netmon feeds your actual bandwidth metrics to an AI model, generating hilariously sarcastic, cynical status updates blaming your local network freeloaders or questionable internet provider for any sudden slowdowns. It is the perfect, highly entertaining way to keep tabs on your home connection.

📰 https://news.ycombinator.com/item?id=49012930

🆔 @hackernewsgithubprojects

📦 auto-index/autoindex

AutoIndex

AutoIndex is the automated development tool that uses collaborative AI agents to write customized text-processing programs. Instead of forcing you to manually tweak how your search engine chunks, parses, and indexes giant documents, this system pairs an Analysis Agent with a Code Agent to automatically optimize retrieval. The Analysis Agent studies retrieval failures to figure out where the search engine is slipping up, while the Code Agent writes and test-runs Python programs to fix those exact weak points. It is like having a tireless engineer run experiments on your search index until it is perfect. You should definitely check out AutoIndex on GitHub to see automated code generation tackle search...

🆔 @hackernewsgithubprojects

📦 nuemaan/skewadam

SkewAdam

An open-source optimizer called skewadam can train massive mixture-of-experts artificial intelligence models on a single graphics card by cutting optimizer memory usage by over ninety-seven percent. Traditional training methods spend massive amounts of memory tracking optimizer details uniformly across every parameter. Instead of treating the model like a single giant block, this tool smartly skews its memory budget. It keeps full tracking details for the frequently visited core parts of the model, but drops unnecessary buffers for the sparse expert layers. This simple shift shrinks peak memory demands so dramatically that a huge six-point-eight billion parameter model can train smoothly on a single forty-gigabyte graphics card, delivering top-tier performance without...

📰 https://news.ycombinator.com/item?id=49002895

🆔 @hackernewsgithubprojects

📦 medlartea/refertrack

ReferTrack

Track target objects with robots using simple natural language commands. Instead of getting tangled up in complex visual processing, this clever system first points out the target you describe in words, and then smoothly tracks it over time. It uses special indicator tokens to feed past tracking data right back into the video history, so the robot never loses its train of thought even when things get blocked or the camera shakes. It is incredibly good at helping humanoid and legged robots follow objects in the real world. Give it a look if you want your robots to actually follow directions.

🆔 @hackernewsgithubprojects

📦 eyeline-labs/go-with-the-track

Go-with-the-Track

Go-with-the-Track is the video generation framework that uses point-tracking to perfectly composite reference images and control motion. Instead of treating point-tracks as simple paths across a video, this project anchors those points directly to your reference images. By establishing exact, frame-by-frame coordinate matches, it gives you incredible control over how objects move and blend. You can easily feed it custom animated meshes from Blender or extract precise trajectories from real footage. This makes it incredibly easy to stylize existing videos or composite completely new elements with realistic, physical consistency. It is a brilliant way to bring complex video editing ideas to life.

🆔 @hackernewsgithubprojects

📦 codeswithroh/tastemaker

Tastemaker

Stop letting AI turn your designs into generic purple gradients. You can now use tastemaker to lock down a genuine design system before your coding agent writes a single line of code. Instead of feeding you a boring menu of pre-selected colors, this clever tool analyzes your project's mood to generate a completely fresh, mathematically contrast-compliant color palette on the spot. It reads real pixels from your reference images, builds custom logos, grabs beautiful illustrations, and even wires up smooth animations automatically. Best of all, it remembers your taste profile locally so your next project gets even better.

🆔 @hackernewsgithubprojects