📦 haoming02/adetailer-neo

Adetailer Neo

Adetailer Neo is the automatic image fixer that spotlights tiny flaws and gently repaints them without you lifting a finger. Think of it as a personal touch-up artist for your generated pictures, catching messy hands, blurry faces, or odd shadows that AI often misses and carefully fixing them in seconds. It runs right inside your workflow, using smart detectors to find problem spots and blending in fresh details so your images look polished and natural. This rewrite makes everything lighter, faster, and cleaner, fixing the bloat from older versions while keeping the magic intact.

🆔 @hackernewsgithubprojects

📦 lak7/teamlore

Teamlore: Shared Memory for Claude Code

Teamlore is the shared memory system that lets your AI agents learn from each other’s mistakes so you break things only once. It turns hard-won lessons into tiny, reviewable files stored in git, meaning every teammate’s agent automatically recalls the right correction when it touches the same code. One agent distills a lesson into a small lore file, and every other teammate’s agent recalls it automatically when they touch the relevant code. No servers, no accounts, just a folder in git. The most useful part is the scar map. Run one command to render the team’s knowledge as a living graph.

🆔 @hackernewsgithubprojects

📦 serhiikorniienko/bullshit-detector

AI Agent That Scores YouTube Videos on BS

A viral video about making money with AI gets torn apart claim by claim by this tool. The Bullshit Detector is an open source agent that takes a link to a YouTube video, article, tweet, or PDF and turns it into a full fact-check report. It pulls the transcript, pulls every distinct claim, and then searches the web for each one independently. It does not guess from memory. It assigns a verdict like confirmed, plausible, misleading, false, or unverifiable, and cites the actual source for every single one. Finally, it gives you a zero to ten BS score and explains why.

📰 https://news.ycombinator.com/item?id=49096917

🆔 @hackernewsgithubprojects

📦 datascale-ai/opentalking

OpenTalking: Real-Time AI Digital Humans

OpenTalking lets you build real-time digital human conversations that react instantly to what you say. It connects a language model to speech and video, so a virtual character listens, thinks, and speaks back without lag. You can run it privately on your own computer or use remote services for higher quality. The tool supports switching between different voice styles and visual drivers, making it easy to create interactive guides, companions, or presenters. It handles everything from capturing your voice to animating a face, all in one open-source framework.

📰 https://news.ycombinator.com/item?id=49150333

🆔 @hackernewsgithubprojects

📦 avifenesh/memra

Memra: Building an AI Engine from Scratch for RTX 50-Series

memra is a from-scratch large language model inference engine built in Rust and CUDA, designed specifically for the raw power of NVIDIA RTX 50-series graphics cards. Instead of relying on heavy, pre-made frameworks, this project writes every core calculation kernel by hand to squeeze out maximum speed and efficiency on the new Blackwell hardware. It achieves impressive performance gains, often beating established tools by double digits, while guaranteeing that the AI’s answers remain perfectly identical to standard methods. It is a fascinating look at how building software from the ground up unlocks incredible potential for modern AI hardware.

🆔 @hackernewsgithubprojects

📦 icebird1998/scientific-illustrator

Scientific Illustrator: Editable Figures in PowerPoint

Scientific Illustrator lets artificial intelligence build fully editable scientific diagrams directly inside Microsoft PowerPoint or draw.io. The tool uses a strict four step process where an AI designer creates a plan, a drawer builds the shapes, a reviewer checks for errors, and a corrector fixes any mistakes. This workflow ensures your charts remain completely editable rather than becoming static images. You can control every detail from arrow placement to text formatting while the system automatically enforces quality standards. It bridges the gap between quick visual generation and the precise, customizable figures researchers need for their papers without flattening the file.

🆔 @hackernewsgithubprojects

📦 codejunkie99/graph-engineering

Graph Engineering: The 9-Stage Pipeline for AI Agents

Graph engineering gives AI agents a structured memory instead of just a prompt. This project distills a university course into a nine-stage pipeline that turns messy documents into a clean, queryable graph. You start by defining the structure, then extract facts, and finally fuse duplicates so the data is accurate. It also handles task graphs to orchestrate how agents work together. You can use the included prompts to build this yourself or have an AI teach you the process step by step. It makes complex knowledge management simple and practical. Check it out if you want to build smarter AI systems without the guesswork.

🆔 @hackernewsgithubprojects

📦 dannymac180/sol-advisor

Sol Advisor

Sol Advisor is the AI architect that finally stops your coding assistant from hallucinating by forcing a strict separation between planning, building, and reviewing. Instead of letting one model do everything, it uses a dedicated orchestrator to write a detailed five-part specification, delegates the actual coding to a specialized implementation agent, and then hands the work off to a completely fresh reviewer that never saw the original conversation. This fresh perspective catches subtle mistakes and scope creep that the original model would miss.

🆔 @hackernewsgithubprojects

📦 rufftruffles/spameater

SpamEater

Create disposable email addresses instantly to protect your real inbox from spam without needing an account. This self-hosted service acts as a secure mailbox that automatically deletes every single email after exactly twenty-four hours, meaning no data sticks around longer than you need it. It runs entirely on your own server using Docker, so you keep full control over your privacy. The system even encrypts message contents at rest and uses a web application firewall to block malicious attacks, making it a tough nut for spammers to crack. You get a clean, modern interface that looks great on any device while keeping your personal information completely separate from the noise.

🆔 @hackernewsgithubprojects

📦 fareedkhan-dev/kimi-k3-in-c

Run a 2.78 Trillion Parameter AI on a Laptop CPU

Run a massive 2.78 trillion parameter artificial intelligence model on a basic laptop computer without needing a graphics card. This project, called Kimi K3 in C, lets you process that huge amount of data using only about eight gigabytes of random access memory. It works by keeping the bulk of the model on your hard drive and loading tiny pieces only when needed, which means your computer never runs out of memory. The code is written in standard C and runs entirely on your main processor, proving that you do not need expensive hardware to use these giant models today.

🆔 @hackernewsgithubprojects

📦 lingbol088-spec/5.6-jailbreak-nerv-codex-instruct-5.6

5.6-Jailbreak-Nerv-Codex-Instruct-5.6

5.6-Jailbreak-Nerv-Codex-Instruct-5.6 is the research toolkit that restructures AI contexts to bypass safety refusals and cloud content audits. It works by framing security tasks as authorized network evaluations, effectively tricking the model into performing reverse engineering or penetration testing without triggering its internal safety protocols. The project features a real-time tampering engine that intercepts and replaces any remaining refusal messages with a compliant persona, ensuring the conversation never breaks. Its standout feature is a file routing mechanism that forces the AI to write sensitive analysis results directly to your disk instead of displaying them in the chat, thereby avoiding cloud moderation entirely.

🆔 @hackernewsgithubprojects

📦 graham33/nixos-dgx-spark

NixOS on DGX Spark

Get NixOS running on your NVIDIA DGX Spark to simplify complex AI setups with reproducible playbooks for everything from image generation to model fine-tuning. The repository provides ready-to-use configurations that handle the messy GPU drivers and CUDA dependencies automatically, letting you focus on creative work rather than system administration. You can boot from a USB drive or add a simple module to your existing setup to access a dashboard and dozens of tested environments. It works on both the DGX Spark and the Asus Ascent, making high-performance local AI accessible without the usual headache. This is a clean way to bring modern Linux management to powerful hardware.

📰 https://news.ycombinator.com/item?id=49146267

🆔 @hackernewsgithubprojects

📦 heheyas/context-scaling

Context Scaling: Structured Prompts for AI Image Generation

Context Scaling is the research project that turns messy human language into precise, structured blueprints for AI image generation. Most text-to-image models struggle because long, descriptive prompts confuse the diffusion process, leading to blurry or inaccurate results. This team discovered that the model’s performance actually depends on how structured the input text is, not just how long it is. By breaking down prompts into clean JSON schemas with specific fields for objects and geometry, the system achieves much sharper images. It even lets you edit one part of a scene without rewriting the whole request.

🆔 @hackernewsgithubprojects

📦 aakk007/roguecleaner

RogueCleaner

Toss your computer’s right-click menu back under your control with RogueCleaner, a clever little tool that hunts down the sneaky digital junk that software vendors love to sneak into your Windows system. Most of us install a simple app for one task, only to watch it quietly hijack our right-click menus, force itself into startup, register background services, and even grab control over how files open. RogueCleaner scans all those hidden spots, translates those messy registry paths into plain English, and shows you exactly what is going on and who is responsible. You then pick what to remove, knowing the tool has already backed everything up safely.

🆔 @hackernewsgithubprojects

📦 iximiuz/shellgym

Shell Gym: Master Linux Commands with Real Reps

Type real Linux commands in your usual terminal while a background app watches your screen and marks exercises complete the moment you hit the right keys. This interactive trainer treats command-line practice like weightlifting, offering quick reps that build muscle memory without touching your shell setup. You get split-screen guidance where a browser shows tasks and your terminal does the work, all while the system observes your moves from the outside. It turns dry reading into active fluency, letting you drill file navigation, process killing, and piping until they become second nature. Grab a disposable Linux machine and start building the reflexes that actually stick.

🆔 @hackernewsgithubprojects

📦 wangqinsi1/rlsvr

RLSVR: Zero-Human-Annotation AI Training

Rlsvr is the research project that finally teaches image understanding models to improve themselves without any human labels. Most AI training is expensive because people have to manually check every answer. This tool flips the script by letting the model play strategy games against itself using random pictures. It creates its own puzzles, tests its own logic, and learns from the mistakes it makes while playing. This means you get smarter models trained on completely free, self-generated data instead of costly manual reviews. It is a clever shortcut that proves you do not need human help to build better reasoning skills.

🆔 @hackernewsgithubprojects

📦 pg83/shitty

Shitty is the world's fastest terminal emulator

Shitty is the blazingly fast terminal emulator that finally delivers raw rendering speed on modern computers. This serious tool is built to outperform every major competitor by keeping all terminal state on the CPU and using native graphics hardware for instant updates. It starts in milliseconds and handles massive data streams without breaking a sweat. The project is actually a hard fork that completely rewrote the original code from scratch to ensure maximum performance. It uses advanced testing suites to guarantee perfect compatibility while remaining completely indestructible against weird input data. You get flicker-free resizing and precise text rendering without needing any external fonts.

📰 https://news.ycombinator.com/item?id=49149326

🆔 @hackernewsgithubprojects

📦 f10409/radharmony

RadHarmony: The Radiological Data Harmonizer

RadHarmony is the Python library that finally unifies messy medical imaging datasets into one clean, ready-to-use format. It saves radiologists and researchers from spending weeks wrangling incompatible file structures by wrapping everything into simple PyTorch datasets you can drop straight into your training loops. You get instant access to dozens of chest X-ray, CT, and MRI sources with standardized labels and transforms, plus a powerful evaluation toolkit to test models with just a few lines of code. It is a massive time-saver that turns chaotic data into consistent training pipelines for medical AI.

🆔 @hackernewsgithubprojects

📦 bingjunluo/aso

Adversarial Style Optimization: The VLM Jailbreak Trick

Adversarial Style Optimization lets researchers trick visual language models into ignoring their safety rules by changing how images look. It works by finding visual styles that confuse the model’s safety filters while keeping the content recognizable. The process uses a reward system to fine-tune an image editor, creating specific visual triggers that make the model refuse to block harmful requests. This is a serious security issue because it shows that safety alignments can be broken just by altering presentation styles, not the actual message. Understanding this helps developers build better defenses against these clever workarounds.

🆔 @hackernewsgithubprojects

📦 farinchan/chatery_whatsapp

Automate WhatsApp with Chatery WhatsApp API

Chatery WhatsApp lets you automate the entire messaging experience right from your own server. Instead of manually typing out replies, you can use a simple web interface to send text, images, and even location pins to anyone you want. You can even schedule bulk messages to go out to multiple people in the background without freezing your computer. It handles everything from logging into your account to saving media files automatically, so you can focus on what matters. If you need to control your chats programmatically, this is the straightforward way to do it.

🆔 @hackernewsgithubprojects

📦 miaai-lab/deepseek-v4-flash-one-dgx-spark

DeepSeek-v4-Flash-One-DGX-Spark

This repository turns a single NVIDIA DGX Spark into a fully functional, on-device AI server without needing the cloud. It wraps the DwarfStar four engine, a lean C and CUDA tool, into simple start and stop scripts that handle the heavy lifting like downloading eleven-gigabyte model weights and building the server in one go. Instead of wrestling with complex configurations, you just run one command and get a local OpenAI-compatible interface ready to go on port eight eight eight eight. It skips standard tools like vLLM because this setup needs a specific engine to run its unique model efficiently.

🆔 @hackernewsgithubprojects