📦 elmma/mllm-reroute

Boost Vision-Language Models with Recoverable Token Routing

Ever wonder how to make large vision-language models more efficient without sacrificing their performance? This repository offers a clever solution called Reroute, which changes how visual tokens are processed. Instead of permanently deleting tokens, which is common in many methods, it uses a smart routing technique that keeps unselected tokens in the residual stream, allowing them to be reconsidered later. This approach helps maintain accuracy while improving overall model efficiency. Whether you are using LLaVA or Qwen models, this codebase provides tools to experiment with these routing strategies and profile performance, giving you better control over your AI.

🆔 @hackernewsgithubprojects

📦 mojixu/trustmargin

Stop Hallucinating: Use TrustMargin to Arbitrate Your RAG Pipeline

Are your language models struggling to decide when to trust their own memory versus retrieved documents? TrustMargin is a training-free solution that acts as a clever judge for your retrieval-augmented generation pipelines. By evaluating a parametric-prior margin and an evidence-binding margin, it dynamically chooses between a direct answer and a retrieval-augmented one for every question. Because it requires no weight updates and relies on simple inference-time logic, it is a lightweight way to boost accuracy without complex retraining. If you want a more reliable way to manage model knowledge and external evidence, this project provides the essential implementation to start.

🆔 @hackernewsgithubprojects

📦 sifter-ai/sifter

Stop RAG. Start Sifting: Transform Unstructured Docs into Data

Are your documents just a dark database sitting in folders? Stop relying on retrieval systems that guess at answers and start using Sifter. This developer-first, open-source engine turns your messy invoices, contracts, and receipts into a structured, queryable database. Instead of searching for text chunks, Sifter extracts typed fields and performs real database-style aggregations that are exact and reproducible. With native Python and TypeScript SDKs, an MCP server for your AI agents, and built-in dashboard generation, Sifter lets you treat your documents like a reliable data source. Take control of your data and turn chaos into clarity today.

🆔 @hackernewsgithubprojects

📦 iamsrikanthnani/pluely

Pluely: The Ultimate Stealth AI Assistant for Meetings

Stop struggling in meetings and get real-time help without anyone knowing. Meet Pluely, an ultra-lightweight, open-source AI assistant designed for complete stealth. Built for performance, this ten-megabyte tool sits as a transparent overlay on your screen, remaining invisible during video calls and screen shares. Whether you need to transcribe system audio, analyze screenshots for instant insights, or use voice input for hands-free queries, Pluely keeps you prepared. With customizable keyboard shortcuts and support for multiple AI providers, it is the perfect privacy-focused companion for any professional. Upgrade your workflow and stay ahead in every conversation today.

🆔 @hackernewsgithubprojects

📦 quantaalpha/quantaalpha

Discover Alpha Factors Effortlessly with AI

Struggling to find the right alpha factors for your quantitative trading strategy? Meet QuantaAlpha, an innovative framework that uses large language models and evolutionary strategies to automate the entire process. Instead of manually searching, you simply describe your research direction, and the system intelligently mines, evolves, and validates factors through self-improving trajectories. It even includes a clean web dashboard for visual monitoring and independent backtesting, taking the complexity out of factor discovery. Whether you are building new models or testing hypotheses, this tool streamlines your research workflow. Give it a try to transform how you approach quantitative finance today.

🆔 @hackernewsgithubprojects

📦 borhen68/tokentamer

Slash Your LLM Costs by 80% with TokenTamer

Are your AI coding agents burning through your budget by constantly resending massive, redundant codebases? Meet TokenTamer, a smart middleware proxy designed to slash your LLM API costs by fifty to eighty percent. By sitting between your coding assistant and the model, it intelligently intercepts payloads to perform real-time AST-based compression. It replaces bloated background code with structural skeletons while keeping your active work intact, and even leverages Anthropic prompt caching to optimize long-lived sessions. Stop paying for context your AI does not actually need to see. Give your wallet a break and start optimizing your developer workflow today.

📰 https://news.ycombinator.com/item?id=48458633

🆔 @hackernewsgithubprojects

📦 braveopotato/fcksignups

Stop Signing Up! Use These Open-Source Browser Tools Instantly

Are you tired of endless registration walls just to use a simple online tool? Meet a curated directory of open-source, browser-based utilities that require absolutely no signups, tracking, or email addresses. Whether you need help with design, development, writing, or data analysis, this project provides a clean, searchable, and responsive interface to find exactly what you need without the bloat. Built with React and TypeScript, this collection prioritizes user privacy and simplicity above all else. Just open the page, search by category or keyword, and get straight to work. Stop wasting time on accounts and start building today.

🆔 @hackernewsgithubprojects

📦 ninoxai/nightwatch

Stop Alert Fatigue: Meet the AI SRE That Investigates Your Outages

Are you tired of waking up at 3am to a flood of alerts that leave you guessing what actually broke? Meet the open-source AI SRE that finally brings clarity to your monitoring. It works as a local-first layer over your existing systems to automatically cluster alert storms into single, actionable incidents and filter out the noise. Most importantly, it features a smart investigator agent that gathers evidence from your live environments to form root-cause hypotheses, all while keeping your data safe and read-only. It never touches your production systems, giving you the control to approve every fix yourself.

📰 https://news.ycombinator.com/item?id=48438180

🆔 @hackernewsgithubprojects

📦 openclaw/crabbox

Crabbox: Instant Cloud Testing for Developers

Stop waiting on sluggish CI pipelines and manage your testing environment with total control. Crabbox is a powerful control plane designed for developers and AI agents to lease managed cloud capacity for remote software testing and execution. It streamlines your workflow by syncing your local code changes to a remote machine, executing tests, and streaming the output back in real time. With support for major cloud providers and local sandboxes, you get cloud-grade compute without the headache of manual setup. Use it to warm up boxes, run commands, and even capture failure details effortlessly. Take command of your remote development today.

🆔 @hackernewsgithubprojects

📦 avafly/paddleocr-ncnn-cpp

Run PaddleOCR Fast with NCNN in C++

Looking to boost your text recognition performance? This C++ project provides a high-performance implementation of PaddleOCR models using the NCNN inference framework. By leveraging NCNN, it enables efficient optical character recognition directly on your own hardware, supporting multiple versions from v3 up to v6. Whether you need fast text detection or accurate recognition, this implementation gives you the tools to run these powerful models with optimized latency. It is a fantastic resource for developers wanting to integrate robust OCR capabilities into C++ applications, offering a lightweight and flexible way to process documents or images quickly.

🆔 @hackernewsgithubprojects

📦 petipoua/petiglyph

Turn Any Video into a Custom Font with Petiglyph

Ever wanted to turn your favorite images or even videos into a custom font? Meet Petiglyph, a powerful terminal tool that lets you transform media files into monochrome font glyphs effortlessly. Whether you are creating static icons from PNGs and SVGs or building animated glyphs from GIFs and MP4s, this project handles the heavy lifting right from your terminal. It offers a smooth interactive interface for managing your font projects and even generates installation-ready files. It is a fantastic way to create unique assets for your own apps or designs. Try it out to start crafting your custom typography today.

📰 https://news.ycombinator.com/item?id=48476626

🆔 @hackernewsgithubprojects

📦 kunchenguid/gnhf

Automate Your Coding Overnight with GNHF

Ever wish you could keep your coding projects moving while you sleep? Meet GNHF, a powerful orchestrator that turns your AI agents into autonomous, overnight workers. Instead of waking up to a stagnant codebase, you get a branch filled with clean, documented changes and a full summary of everything the agent accomplished. It is agent-agnostic, meaning it works seamlessly with tools like Claude Code, GitHub Copilot, and more. By handling incremental commits and automatic error recovery, it ensures progress never stops. Stop wasting downtime and let your agents handle the heavy lifting while you rest.

🆔 @hackernewsgithubprojects

📦 levy-street/world-of-claudecraft

Build and Play Your Own Vanilla-Style MMO

Ever wanted to host your own classic-style micro-MMO? This project delivers a fully playable, vanilla-flavored multiplayer game that you can run right in your browser or host as a persistent server. It features a deterministic simulation engine that keeps the world synchronized, complete with character persistence, trading, dueling, and party mechanics. You can explore a vast, procedural world with nine distinct character classes, challenging dungeons, and deep combat systems. Whether you want to develop your own world or simply jump in for an offline adventure, this project provides a polished, open-source foundation for your next gaming obsession.

🆔 @hackernewsgithubprojects

📦 lishix520/academic-paper-skills

Master Academic Writing with Claude Code

Struggling to turn your research ideas into a polished manuscript? This framework offers a systematic way to plan and write academic papers using Claude Code. By separating the process into a strategist phase for planning and a composer phase for writing, it ensures your work meets high standards through built-in quality checkpoints and reviewer simulation. It helps you identify research gaps, maintain structural consistency, and prepare submission-ready drafts with Python-based validation tools. Whether you are a student or an independent researcher, these skills provide a professional workflow to elevate your academic writing. Start organizing your research today and finish stronger.

🆔 @hackernewsgithubprojects

📦 jjang-ai/mlxstudio

Run Local AI on Mac with MLX Studio

Stop relying on the cloud and start running powerful AI models directly on your Mac. MLX Studio is a native macOS desktop application that lets you run LLMs, vision models, and image generators entirely offline. It provides a complete environment for chatting, coding, and image editing while keeping all your data private. You can easily download models from Hugging Face, utilize advanced adaptive mixed-precision quantization for better performance, and even launch an OpenAI-compatible API server for your local workflows. If you want a smooth, private, and high-performance way to explore local AI, this tool is your perfect companion.

🆔 @hackernewsgithubprojects

📦 runnin4ik/dpi-detector

Is Your Internet Being Censored? Test It Now!

Ever wonder if your internet connection is being restricted by deep packet inspection? This powerful tool helps you test and analyze traffic censorship right from your own device. It identifies specific connection drops, checks for DNS tampering, and categorizes various blocking methods like TCP resets or TLS issues. Whether you are troubleshooting access to specific sites or simply auditing your network security, this project provides a clear view into what is happening behind the scenes. It is a fantastic resource for anyone interested in network diagnostics, offering a straightforward way to verify your connection status and stay informed.

🆔 @hackernewsgithubprojects

📦 baairon/soundcli

Take Control of Your Music Library with SoundCLI

Stop relying on streaming services and start owning your music collection. SoundCLI is a powerful terminal based tool that lets you download your favorite tracks from YouTube, SoundCloud, and Spotify directly to your computer for offline listening. It automatically handles artist details and album artwork, organizing everything into your own local files. If a download is interrupted, it intelligently resumes exactly where it left off, and it never downloads the same song twice. Everything is managed through an intuitive, keyboard driven interface, giving you a fast and private way to curate your library. Reclaim your music today.

🆔 @hackernewsgithubprojects

📦 danmcinerney/architect-loop

Automate Coding with a Multi-Agent Architect Loop

Tired of AI agents drifting off-course or hallucinating their own success? Architect Loop solves this by separating planning from execution using a smart cross-vendor approach. It tasks Claude Fable to act as a rigorous architect that sets specs and enforces strict gates, while GPT-5.5 Codex handles the heavy lifting as a builder. By using the repository itself as the only source of truth, it ensures all work is verified, isolated, and truly mergeable before integration. Whether you are building features or performing deep research, this setup keeps your agents focused, disciplined, and accountable. Try this robust pattern today.

📰 https://news.ycombinator.com/item?id=48509133

🆔 @hackernewsgithubprojects

📦 haddock-development/claude-reflect-system

Stop Repeating Yourself! Use Claude Reflect to Make Claude Code Learn Permanently

Are you tired of correcting your AI coding assistant over and over again? The Reflect system for Claude Code finally solves this by enabling permanent, self-improving learning directly on your machine. Whenever you provide a correction or feedback, Reflect captures that interaction to update your AI skills automatically. It uses a smart confidence-based system to detect your patterns and preferences, ensuring Claude remembers your specific tools, project structures, and coding standards. With built-in features like Git version control, interactive review, and timestamped backups, you stay in full control of every update. Start teaching your AI today and enjoy a truly personalized workflow.

🆔 @hackernewsgithubprojects

📦 crmne/kamal-backup

Automate Your Rails Backups with Kamal

If you are managing a Rails application with Kamal and find yourself juggling complex backup scripts, you need to check out kamal-backup. This repository provides a simple, native way to handle automated, encrypted backups for your database and file-backed storage as a standard Kamal accessory. It solves the headache of fragmented backup maintenance by wrapping everything into a single gem and Docker image. You get scheduled snapshots, support for PostgreSQL, MySQL, and SQLite, plus built-in tools for restore drills and generating security review evidence. Stop worrying about data safety and start automating your production backups with confidence today.

📰 https://news.ycombinator.com/item?id=48051178

🆔 @hackernewsgithubprojects

📦 kouhxp/textsnap

Turn Any Image or Webpage Into Text Instantly, Fully Offline!

Tired of manual transcription? Meet Textsnap, the ultimate tool for turning images, screenshots, or entire webpages into editable plaintext with just one command. Whether it is a local photo, a direct image link, or an online article, Textsnap processes everything locally on your own CPU without needing a GPU or the cloud. It leverages an advanced vision-language model to accurately extract structured content, keeping your data completely private and offline after the initial setup. Best of all, it works directly with your clipboard for a seamless workflow. Simplify your digital life and start capturing text effortlessly today.

📰 https://news.ycombinator.com/item?id=48410841

🆔 @hackernewsgithubprojects