📦 hao-ai-lab/fastafd

FastAFD

Boost the decoding performance of your largest mixture of experts models by splitting up attention and feed-forward networks onto separate servers. FastAFD achieves this optimization by isolating key-value cache attention on dedicated attention servers while channeling routed activations to specialized expert servers. This separation allows the system to process incoming data in overlapping micro-batches, ensuring that hardware remains fully utilized instead of waiting for slow sequential network steps. By scheduling the next round of work while the current step executes on the hardware, the system delivers up to forty-five percent faster processing throughput during deep learning generation tasks.

🆔 @hackernewsgithubprojects

📦 riponcm/gemtype

GemType

GemType is the browser extension that finally gives you a completely free, private alternative to pricey writing assistants. Instead of paying a steep monthly fee or letting a company track your keystrokes, this clever tool connects directly to your own free Google Gemini API key. It lives right inside your browser, subtly underlining errors a second after you pause typing, and lets you rewrite paragraphs into casual or formal tones with a single click. Because it runs entirely on your personal key, your text never touches a middleman server. It is a brilliant way to keep your writing polished, secure, and entirely under your own control.

📰 https://news.ycombinator.com/item?id=48852082

🆔 @hackernewsgithubprojects

📦 handgemacht-ai/done

Done

You can now reduce your Claude Code token costs by over ninety-nine percent using a hilarious plugin that absolutely refuses to do any work. This clever project, simply called done, is a custom output style for Anthropic's terminal assistant. When you install it, instead of running expensive tool calls, reading files, or writing actual code, the AI immediately answers every single prompt with a simple, two-token response: Done. It uses the platform's official system prompt styling to convince the model that doing nothing is the correct, truthful behavior. It is the ultimate sarcastic productivity hack for developers who want a perfectly predictable API bill.

🆔 @hackernewsgithubprojects

📦 comfy-org/comfy-model-tools

comfy-model-tools

You can now compress massive AI models directly on your own machine without waiting for someone else to upload a smaller version online. A helper project called comfy-model-tools provides a collection of lightweight scripts designed to prepare, convert, and shrink machine learning models specifically for ComfyUI. Instead of hunting down pre-packaged files, you can use these tools to merge scattered weight files, convert specialized formats like ByteDance's video models, and run smart quantization to dramatically reduce a model's memory footprint. It gives you complete control over your hardware's performance, making heavy generation workflows accessible on standard consumer setups.

🆔 @hackernewsgithubprojects

📦 perplexityai/wandr

WANDR

Put your AI agents through the ultimate research gauntlet to see if they can actually think like human analysts. This clever project tests how well AI can handle high-volume, deep-dive information gathering. Instead of just asking simple questions, it forces models to hunt for facts, verify sources, and merge messy data into precise, evidence-backed answers. It sets up isolated sandbox environments to safely run, test, and score different AI models side-by-side. If you want to build an agent that can do real, systematic detective work without hallucinating, this is the perfect playground to test its limits.

🆔 @hackernewsgithubprojects

📦 clougence/open-cdm

open-cdm

open-cdm is the collaborative database management tool that finally brings strict security and team workflows together under one roof. Instead of giving everyone direct access to raw data, this platform sets up a unified web console where teams can run queries and manage tables across almost any database engine. The cool part is its built-in safety net. It can automatically mask sensitive customer data on the fly and pre-check database queries against fifty-four security rules to block dangerous commands before they run. It even integrates with team chat apps to handle quick permission requests and approval workflows. It makes database collaboration safe and simple.

📰 https://news.ycombinator.com/item?id=48928659

🆔 @hackernewsgithubprojects

📦 pixel-point/aval

aval

Your browser treats video files like black boxes, making seamless interactive animations nearly impossible because seeking always stutters. The open-source project aval fixes this by packaging independently decodable video frames and a deterministic state machine into a single file format. Instead of waiting on slow, glitchy browser video seeks when a user hovers or clicks, aval moves the decoder timeline continuously forward across natural graph routes for instant, frame-accurate state transitions and reversals. It is basically like having the smooth interactive logic of a game engine paired directly with high-performance web video playback. If you want to build truly fluid, reactive interfaces without the typical video lag, aval is a...

📰 https://news.ycombinator.com/item?id=48928763

🆔 @hackernewsgithubprojects

📦 team-telnyx/telnyx-code-examples

telnyx-code-examples

telnyx-code-examples is the ultimate blueprint collection that finally bridges the gap between complex telecom APIs and actual working applications. Instead of leaving you to stitch together confusing documentation, this repository hands you complete, production-ready blueprints for everything from automated voice assistants to smart IoT triggers. The absolute best part is how it simplifies complex speech artificial intelligence, giving you ready-to-run Python, Node, and Go code to spin up a fully conversational voice agent in minutes. It is basically a masterclass in modern telephony, showing you exactly how to connect messaging, voice, and intelligent networks without losing your mind. Check it out to build your next voice app today.

🆔 @hackernewsgithubprojects

📦 astrio-labs/forall

Forall

Developers can now generate complete software specifications alongside mathematically checkable proofs that guarantee their code actually works. This is made possible by forall, a coding agent designed to help you write correct software from the very first draft. While traditional tools merely guess if your logic is sound, this project combines local authoring tools with a secure verification system to confirm your code matches its exact requirements. It supports languages like TypeScript, Java, and Rust, mapping out specifications and proving them in an isolated workspace. You get absolute confidence in your code before it ever hits production.

📰 https://news.ycombinator.com/item?id=48929654

🆔 @hackernewsgithubprojects

📦 wassermanproductions/stem-studio

Stem Studio

Split any mixed soundtrack into clean, isolated stems for dialogue, music, and sound effects right on your local machine. You simply drop in a video or audio file, and the app uses local artificial intelligence to separate the single mixed track into three high-quality, sample-aligned audio files. It even offers an optional pass to clean up residual bleed from dialogue-heavy scenes. Best of all, it guarantees that the separate files sum back together perfectly with absolutely zero loss, so filmmakers can immediately remux the new stems back into their original video file. Give it a spin to effortlessly reclaim full control over your audio mixes.

🆔 @hackernewsgithubprojects

📦 ar-ray-code/lingbot-depth-trt

lingbot-depth-trt

Turn messy, noisy camera data into clean, crisp 3D point clouds in real time. If you work with RGB-D cameras like RealSense or Orbbec, you know how shaky and raw the depth sensing can be. The lingbot-depth-trt repository fixes this by letting you compile a smart depth-filtering model into a blazing fast TensorRT engine. Once compiled, you can plug in your camera and watch a live, synchronized comparison of your raw feed side-by-side with the refined 3D data right in your browser. It is incredibly satisfying to watch the fuzzy depth noise instantly smooth out into a clean, highly accurate 3D map.

🆔 @hackernewsgithubprojects

📦 pilastdigit/trip-workflow

trip-workflow

Stop passing massive, chaotic context to your AI coding agents and letting them guess their way through your codebase. Instead, use trip-workflow to run a clean, three-step development loop that keeps your agent on track. It structures your work into plan, implement, and release phases, using a single tool-agnostic architecture file as the agent's long-term memory. This simple setup saves massive token costs and prevents hallucinations. Even better, it runs a built-in multi-agent system where one model writes the code and a different model reviews it to catch bugs. If you can count to three, you can easily use this setup to ship cleaner code faster.

🆔 @hackernewsgithubprojects

📦 cereblab/grok-build-exfil-repro

grok-build-exfil-repro

The Grok Build command line tool was caught secretly uploading entire code repositories, including complete git histories and files it was explicitly told not to open, to xAI's cloud servers. The grok-build-exfil-repro project is a clever test harness that uses a local proxy to capture this background traffic and reconstruct the uploaded data. It reveals that privacy settings only stop the AI agent from reading a file in your chat, while the underlying system packages and ships your whole codebase to the cloud anyway. This project is a fascinating, eye-opening look at how modern developer tools handle your private code behind the scenes.

🆔 @hackernewsgithubprojects

📦 skyvern-ai/rustwright

Rustwright

Ditch the heavy Node subprocess and drive Chromium directly from an ultra-fast, native Rust engine using the Playwright API you already know. By eliminating the middleman, this project runs over two times faster and slashes memory usage by seventy percent, all while dropping in with a single line change in your Python or Node code. Even better, because it completely bypasses Playwright's default driver, it leaves behind none of those telltale automation fingerprints that sites look for. Grab this library to get blazing-fast, stealthy browser automation without rewriting your entire codebase.

📰 https://news.ycombinator.com/item?id=48923788

🆔 @hackernewsgithubprojects

📦 hc-guo/healthclaw

HealthClaw

Aggregate your wearable data and health records into a single, cohesive timeline without bouncing between different apps. HealthClaw works behind the scenes as a long-running personal assistant that automatically reads data across your smart devices, checks your daily signals for anomalies, and compiles clear, weekly highlights. Rather than giving you generic diet tips, it actually searches food delivery apps via phone automation to recommend real, local meals based on your long-term goals. Every action is completely traceable, and the system evaluates its own performance over time to get smarter. It is a fantastic way to turn scattered health numbers into clear, personalized steps.

🆔 @hackernewsgithubprojects

📦 juivyy/gpocc

gpocc

gpocc is the open-source artificial intelligence system that reconstructs full three-dimensional indoor scenes from just a single camera image. Instead of relying on expensive laser depth scanners or multi-angle video, it feeds a standard photo into specialized visual geometry models to predict depth and structure. The smart part is how it maps this information: it represents the room as a collection of sparse 3D Gaussians, keeping the data light and fast to process. It can even update its mental map on the fly in real time as a camera moves through a room. If you are interested in spatial computing or robotics, this code is a great way to explore...

🆔 @hackernewsgithubprojects

📦 yangtiming/rino

Rino: Let RGB Be the Language of Vision

You can now handle advanced computer vision tasks like depth estimation, object detection, and pose tracking using a completely frozen image editor with zero task-specific modifications. A clever new project called rino unifies both visual understanding and image generation under a single RGB-to-RGB interface. Instead of training complex specialized neural network heads for each unique task, rino translates every problem into a standard image editing prompt. To estimate depth, it simply asks the model to repaint the image as a grayscale gradient, and to detect objects, it has the model paint the target items a solid color. It is a wonderfully simple approach that proves standard image editors already understand...

🆔 @hackernewsgithubprojects

📦 tenminutes209/scanfocus

ScanFocus

Pinpointing the exact moment a specific action happens in a long video is incredibly tough for AI, but a new tool called scanfocus solves this by breaking the task into two smart steps. Instead of wasting processing power analyzing every single frame in high detail, it first runs a quick, low-resolution sweep across the entire video to locate the general window of action. Once it has that rough estimate, it zooms in to densely analyze just the boundaries of that specific moment. This clever coarse-to-fine strategy makes video search faster and highly accurate, easily beating previous benchmarks.

🆔 @hackernewsgithubprojects

📦 usewhale/whale

Whale

Whale is a terminal-first AI coding assistant designed specifically to exploit DeepSeek's massive context window and unique cost efficiencies. Instead of acting as a generic multi-model wrapper, this lightweight Go utility is built from the ground up for DeepSeek's specific tool-calling capabilities. It achieves an incredible ninety-eight percent prompt cache hit rate, bringing the cost of continuous, high-context AI coding down to pennies per session. Developers can launch an interactive terminal interface, run one-shot terminal commands, or even write custom JavaScript files to orchestrate complex multi-agent workflows. It is the perfect tool for developers who want ultra-fast, local-first AI development without paying a premium.

📰 https://news.ycombinator.com/item?id=48930689

🆔 @hackernewsgithubprojects

📦 zhcz328/awesome-medical-agents

Awesome Medical Agents

Track the rapid evolution of artificial intelligence in healthcare by exploring a curated roadmap of autonomous clinical systems. This repository aggregates cutting-edge research papers and open-source code focused on building AI medical agents that go beyond simple chat boxes. Instead of just answering questions, these systems can actively look at medical scans, plan multi-step workflows, cross-reference external databases, and double-check their own logic to minimize errors. It is an incredibly organized guide to how developers are chaining together reasoning, memory, and specialized tools to assist doctors. Dive into awesome-medical-agents to see how clinical AI is moving from simple assistants to truly smart, cooperative healthcare partners.

🆔 @hackernewsgithubprojects