📦 weschera/spark-bench

Spark Bench

Spark-bench is the specialized performance benchmark that finally shows you how large language models actually behave when deployed on real hardware. Built specifically for operators of high-performance NVIDIA hardware, it puts models through seventy-four complex scenarios across twelve distinct domains to see how they handle tough, practical tasks. Instead of relying on simple academic questions, this project tests multi-turn agentic workflows, complex tool use, and long-context processing under real cluster serving conditions. It scores models on quality, calibration, and reliability, rather than just raw speed. It is the perfect tool for developers trying to choose the exact model recipe to serve on their local infrastructure.

🆔 @hackernewsgithubprojects

📦 dn9uy3n/modern-red-team-infrastructure

modern-red-team-infrastructure

modern-red-team-infrastructure is the master blueprint that finally maps out a complete offensive security stack from scratch. Instead of hunting down separate tools, this project acts as a central architecture model that organizes everything a professional security team needs to simulate real-world cyber attacks. It brings together advanced command-and-control frameworks, specialized training resources, and automated scanning platforms into one cohesive plan. What makes it incredibly interesting is its forward-looking focus on artificial intelligence, integrating smart agents that can hook directly into development tools to help humans deploy security campaigns. It is a fantastic reference for understanding how modern defensive testing is built.

🆔 @hackernewsgithubprojects

📦 gokulrajaram/productspec

productspec

Define clear software goals before writing a single line of code by using a portable markdown standard that keeps humans and AI agents perfectly aligned. This repository provides productspec, an open format for capturing the strategic what and why of a feature before engineering begins. Instead of losing vital decisions in messy threads or ticketing systems, this tool structures key details like scope, acceptance criteria, and success metrics into a machine-readable format. It even includes specialized skills that guide AI coding agents to build exactly what you intended. It is the perfect way to make sure your original product vision actually survives the handoff to implementation.

📰 https://news.ycombinator.com/item?id=48828135

🆔 @hackernewsgithubprojects

📦 opensensenova/sensenova-vision

SenseNova Vision

Handle virtually any computer vision task with a single AI model instead of swapping between dozens of specialized tools. This clever project turns complex vision challenges like depth estimation, object tracking, and image segmentation into simple text and image generation. Instead of relying on rigid, custom-built pieces of code for every separate task, it feeds instructions to a unified model that responds naturally with text, coordinates, or newly generated masks and depth maps. It is like having a Swiss Army knife for visual data that actually understands what you want. Grab the code and try the web demo to see how easy unified vision can be.

🆔 @hackernewsgithubprojects

📦 snowflake-ai-research/arctic-platform

arctic-platform

During reinforcement learning training for large language models, up to ninety-five percent of the computation is completely wasted on recalculating the exact same prompt over and over. The open-source project arctic-platform solves this exact bottleneck by introducing automatic prompt deduplication during training and inference. Instead of running redundant attention calculations for every single response, it packs unique prompts together, processes them once, and transparently reconstructs the results. This simple optimization drastically reduces memory usage and boosts throughput, especially for tasks with long contexts. It is a practical, lightweight tool that helps developers train smarter models much faster without wasting expensive GPU power.

🆔 @hackernewsgithubprojects

📦 chattocorp/chatto

Chatto

Chatto lets you run a complete, professional chat server on your own hardware using a single binary with an embedded database. Instead of relying on heavy databases, it uses NATS JetStream as its primary storage engine to stream and persist every message, reaction, and voice call event. It is built with a Svelte frontend and a Go backend, giving you a lightning-fast alternative to Slack or Discord that you fully control. You get threaded replies, real-time presence, and full user access management out of the box. It is the ultimate tool for developers who want a self-hosted chat platform that is simple to deploy and incredibly fun to run.

📰 https://news.ycombinator.com/item?id=48863875

🆔 @hackernewsgithubprojects

📦 yifanzhang-pro/hla

HLA

The causal attention mechanism that finally handles complex token interactions in linear time, hla solves the massive memory bottleneck of processing long text. Standard AI models slow down dramatically with long prompts because their memory usage grows quadratically as context expands. While older linear alternatives exist, they often simplify details and lose accuracy. This project introduces a second-order design that keeps memory usage constant and computes updates word by word. It uses clever mathematical shortcuts to run parallel training on modern hardware while perfectly preserving sequential accuracy. It is a major step toward making incredibly long AI conversations fast, cheap, and highly capable.

🆔 @hackernewsgithubprojects

📦 mgechev/skillgrade

skillgrade

You can now run unit tests on AI agents to see if they actually know how to use your custom tools and instructions. A lightweight utility called skillgrade makes this incredibly easy by orchestrating simulated work environments inside Docker containers. It sends instructions to your agent, lets it perform tasks, and then scores the outcome. You can write simple deterministic scripts that verify files or run an automated AI rubric to grade the agent's workflow quality. Best of all, it works locally or in your CI pipeline, and you can preview the results in a clean browser dashboard. It is the perfect tool for proving your agents actually work before...

📰 https://news.ycombinator.com/item?id=48864778

🆔 @hackernewsgithubprojects

📦 esi-bench/esi-bench

esi-bench

esi-bench is the spatial intelligence benchmark that finally treats AI observers as active physical agents. Instead of feeding models a static set of images and asking them to guess what is happening in a room, this project lets them drive. Built on a physics simulator, the system forces models to actively navigate, look under objects, and sequence physical actions to answer complex questions about their surroundings. It turns out that when models can move around to investigate, they easily outperform passive vision, but they still struggle with choosing smart actions and knowing when their information is incomplete. It is a fascinating look at the future of robotic common sense.

🆔 @hackernewsgithubprojects

📦 tencent-hunyuan/hils-attention

HiLS-Attention

Train language models to handle virtually infinite text length without slowing to a crawl. The hils-attention repository introduces a smart way to manage computer memory during long chats or massive document reading. Normally, artificial intelligence gets bogged down because it tries to look at every single word at once. This project solves that by compressing large blocks of text into smaller summaries, letting the model instantly pick only the most relevant sections to focus on. This means you get lightning-fast processing and much better memory extrapolation. Give hils-attention a look to see how it makes ultra-long AI context incredibly efficient.

🆔 @hackernewsgithubprojects

📦 kacper-daftcode/vllm-moet

vLLM-Moet

Serve massive frontier models like GLM-5.2 and DeepSeek-V4-Flash directly on consumer Blackwell graphics cards without running out of memory. While these massive models normally require enterprise data centers, vllm-moet uses custom-written machine code kernels to compress the massive mixture-of-experts layers down to just two bits, while keeping a clever four-bit cache to restore original quality on the fly. When a model is still too big, it seamlessly turns your graphics card into a cache, keeping the active parts on the card and streaming the rest from your system memory or SSD with zero loss in output quality. Grab the code, boot up a model, and run massive AI on your...

🆔 @hackernewsgithubprojects

📦 majidmanzarpour/threejs-game-skills

Three.js Game Skills

Build fully playable, high-quality browser games using artificial intelligence assistants that act as your professional game development team. This repository installs custom skills into agents like Claude Code, allowing them to autonomously write gameplay logic, generate three dimensional models, compose sound effects, design slick user interfaces, and run automated playtests. Instead of writing boilerplate code, you simply tell your helper to construct a game from scratch, and it coordinates specialized workflows to make it happen. It is an incredibly powerful way to bring complex interactive ideas to life in your browser instantly.

🆔 @hackernewsgithubprojects

📦 cpp2rust/cpp2rust

cpp2rust

cpp2rust is the compiler tool that automatically translates old C++ codebases into fully safe Rust. Moving away from C++ usually means rewriting everything by hand to satisfy Rust's strict safety rules, but this tool takes a different path. It parses your existing code, maps out how the data flows, and outputs clean Rust files. To keep things safe without breaking how your code behaves, it uses a clever helper library that manages complex C++ pointers behind the scenes. This gives you a fast, automated way to upgrade legacy software to a modern language without starting from scratch.

📰 https://news.ycombinator.com/item?id=48862086

🆔 @hackernewsgithubprojects

📦 ather-techie/rag-interview-system

rag-interview-system

Preparing for AI developer interviews just got much easier with rag-interview-system, a massive open source handbook featuring over five hundred real-world questions and answers covering forty-one different retrieval-augmented generation architectures. Landing a modern AI job requires knowing more than just basic document search, and this guide walks you through complex setups like knowledge graphs, self-correcting retrieval pipelines, and speculative generation. It goes beyond theoretical concepts to prepare you for realistic system design rounds by analyzing critical production failures like semantic cache leakage and conversational drift. Check it out to level up your engineering skills before your next big interview.

🆔 @hackernewsgithubprojects

📦 utahmobl/opencap-monocular

OpenCap Monocular

You can now extract full 3D human biomechanics and skeletal motion from a single smartphone video. The opencap-monocular repository achieves this by combining 3D pose estimation, camera optimization, and biomechanical modeling. Traditionally, capturing physics-grade joint angles and muscle forces required expensive multi-camera labs, but this project processes a single-camera feed to generate accurate files compatible with OpenSim. It even integrates with an activity classifier to automatically recognize movements like squats or walking and optimize the calculations. It is a massive win for sports science and physical therapy, turning any phone into a portable biomechanics lab.

🆔 @hackernewsgithubprojects

📦 sazardev/goca

Goca

Build perfectly structured Go applications in seconds without writing any boilerplate code. If you are starting a new Go service, this tool scaffolds an entire Clean Architecture setup with one command. Instead of manually creating domain entities, use cases, database repositories, and HTTP or gRPC handlers, you just tell it what fields you need. It generates clean, decoupled layers, hooks up dependency injection, and even writes your integration tests automatically. It even includes safety guards like dry-run previews and automatic file backups so you never accidentally overwrite your work. Grab goca and start writing clean Go code today.

🆔 @hackernewsgithubprojects

📦 rodiun/frugon

Frugon

frugon is the local command-line tool that finally shows you how to slash your artificial intelligence bills without sending your private data anywhere. It runs entirely on your own machine, analyzing your raw API logs to calculate exactly how much cash you would save by swapping out expensive models for cheaper alternatives. You can set up its local proxy shim to capture your traffic, and the tool immediately highlights which simple requests can be routed to budget-friendly models and which complex ones need the premium tier. It is the ultimate way to stop overpaying for token usage and optimize your setup today.

📰 https://news.ycombinator.com/item?id=48816724

🆔 @hackernewsgithubprojects

📦 markfulton/claude-antigravity-agents

claude-antigravity-agents

The claude-antigravity-agents skill is the Claude Code extension that finally lets you run heavy background coding tasks in parallel without burning your subscription tokens. It turns Claude into an active orchestrator that delegates massive jobs like full-repo audits, security reviews, and large refactors to Google's Antigravity CLI. While you and Claude keep building your main features, a secondary model runs the grind work in an isolated background sandbox. Once the sub-agent finishes, Claude automatically runs build checks and reviews the code changes before merging them, giving you a safe, multi-model workflow that multiplies your coding speed.

📰 https://news.ycombinator.com/item?id=48817248

🆔 @hackernewsgithubprojects

📦 gokapso/whatsapp-cloud-inbox

WhatsApp Cloud Inbox

Manage multiple corporate chat accounts from a single dashboard using an elegant web interface. This project handles the official chat network API, giving you a shared interface that mimics the familiar messaging experience. The coolest part is how it automatically solves API limitations, like checking if your last customer interaction was within twenty-four hours and switching to authorized messaging templates if it has been longer. It handles custom interactive buttons, documents, and live media while tracking read receipts across different business numbers. Try running this simple self-hosted setup to keep all your customer chats organized in one place.

🆔 @hackernewsgithubprojects

📦 op7418/guizang-material-illustration

guizang-material-illustration

guizang-material-illustration is an AI assistant skill that creates high-quality explanatory diagrams and material-style charts featuring integrated Chinese text labels. Instead of generating generic decorative images, it translates complex ideas, data, or concept drafts into clear visual guides like flowcharts, mechanism diagrams, and beautiful 3D charts. By focusing on the central illustration layer, it cleanly extracts data from messy screenshots or raw text, packages the visual metadata, and renders readable diagrams perfect for slides, documentation, or social media cards. It is an incredibly clever tool for transforming abstract explanations into instantly understandable visual summaries.

🆔 @hackernewsgithubprojects