Show HN for October 22, 2025

This is a recent addition to Vexlio which I think the HN crowd may find interesting or useful.

TL;DR: easy creation of interactive diagrams, meaning diagrams that have mouse click/hover hooks that you can use to display pop-up content. The end result can be shared with a no-sign-in-required web link.

My thought is that this is useful for system docs, onboarding or user guides, presentations, etc. Anything where there is a high-level view that should remain uncluttered + important metadata or details that still need to be available somewhere.

You can try it out without signing up for anything, just launch the app here (https://app.vexlio.com/), create a shape, select it with the main pointer tool and then click "Add popup" on the context toolbar.

I'd be grateful for any and all feedback!

We built a search engine for art that uses natural language prompts to find real artwork instead of generating art.

We built RuleHunt to search for interesting cellular automata rules using tiktok-style engagement monitoring.

The first rule shown is Conway’s Game of Life. As you scroll, you will see other random rules – the search space is 2^512. Help us find good rule heuristics by starring the ones you like!

On mobile it's a tiktok-like scrolling interface; on desktop it's an interface for targeted rule searching. Starred rules go to a global leaderboard.

GitHub repo: https://github.com/rulehunt/rulehunt

Built this for streaming AI tool calls. LLMs stream function arguments as JSON character-by-character. Most parsers reparse from scratch each time - O(n²) behavior that causes UI lag.

This maintains parsing state, processing only new characters. True O(n) performance that stays imperceptible throughout the entire response.

Ruby gem, MIT licensed. Would love feedback.

Two years after open sourcing Proton, we're releasing v3.0 which brings enterprise-grade streaming capabilities to the open source version: full connectivity, processing, and routing in a single binary with zero dependencies.

Key features:

First vectorized streaming SQL engine in modern C++ with JIT compilation High-throughput, low-latency, high-cardinality processing End-to-end streaming: ETL, joins, aggregation, alerts, and tasks Native connectors: Kafka, Redpanda, Pulsar, ClickHouse, Splunk, Elastic, MongoDB, S3, Iceberg Native Python UDF/UDAF support to support your AI/ML work loads

The same performance we've proven in large enterprise deployments is now available in the community edition.

Would love feedback from anyone working with streaming data or looking for Flink/ksqlDB alternatives.

Hi HN, I'm a solo dev who built this because I kept opening my phone and falling into social media instead of working on myself.

Coach Thyself is a journaling app with structure instead of blank pages: - Daily prompts and community questions to write about - Goal tracking and task organization built in - Optional insights from philosophy/astrology/psychology as reflection triggers - No social features, followers, or performance metrics

I use it every morning now instead of scrolling. It's the motivation that keeps me working on it.

Tech stack: [whatever you used - mention it if interesting] 7-day free trial, $12/month after: https://thyself.coach

Happy to answer questions about the build process or journaling habits. Looking for feedback on what would make this more useful.

Tired of losing focus when waiting for codegen to finish? Now you can enhance your concentration by integrating subway surfers into your workflow.

brew tap evanreilly/subway-surfers brew install subway-surfers Usage subway-surfers

Hi HN! I'm Amber (13) and my dad Raj, and we built Nityasha AI from Guna, India. After my dad's 12 years of failed startups (2012-2023), we created a personal AI assistant that handles email, coding help, research, and planning in one conversational interface.

I started coding at 9 on a 4GB RAM laptop. We failed 8 times before this—coupon sites, freelancing platforms, consulting. Nityasha is different: it uses Thesys generative UI for visual charts, includes Study Mode with Socratic teaching, and integrates everything so you don't need 10 tabs open.

500+ active users now. We just launched Nityasha Connect where businesses can integrate services directly into the AI.

Would love your feedback!

Maktabah Shamilah counterpart: ELKIRTASSE is now being revived, readily to be improved on Qt6 and cmake build system Please help develop, improve it

Show HN: SerenDB – Neon PostgreSQL fork optimized for AI agent workloads

We forked Neon to make database operations faster and safer for AI agents. The goal is to enable instant experimentation with production data and catch prompt injection attacks before they hit your DB. The open-source repo is at https://github.com/serenorg/serendb

The coolest current features are:

1. Time-travel queries: Query your database as it existed at any timestamp. SELECT * FROM orders AS OF TIMESTAMP '2024-01-15 14:30:00'. Essential for debugging agent decisions and auditing what data an agent saw.

2. Scale-To-Zero with pgvector: Built-in vector embeddings that sleep when idle. Pay nothing for dormant agent databases, auto-scale to 16 vCPU when needed.

The coolest features in development are:

1. Prompt injection detection: Fingerprinting and context-aware policies to catch injection attempts before they reach your data. We're extending Postgres's security model to understand AI-specific attack patterns.

2. 100ms branch creation: Clone your entire production database in 100ms (vs Neon's 500ms). Enables branch-per-agent patterns and instant rollback. Great for faster testing of multiple prompt variations on real data.

This is early - we're working on the hosted service now.

Technical folks: curious if anyone else is hitting the "how do I safely test my agent on prod data" problem. Happy to answer questions about our updates to the branching architecture or our proposed context security approaches.

Website: https://serendb.com/

3hours to make, 100% vibe coding inspire by https://news.ycombinator.com/item?id=39159418

Hi HN – over the weekend I hacked together a tiny web tool that detects faces in an image and hides them with a single click. It runs entirely in your browser (no servers, no accounts) and is meant for those moments when you want to share a photo but not share everyone’s identity.

Drag a picture onto blurfaces.org, hit Blur, and it will obscure all the faces it finds. You can toggle individual masks or tweak how strong the blur is. For now it works with still photos only, but it’s been surprisingly handy for family shots and marketplace listings.

I’d love to know: does this scratch an itch for you? What’s missing or annoying? Would you use a tool like this, or do you already have an easier workflow?

Hi HN, back in October 2024 I've launched Buzzd Chat, a Y!M revival project.

I started it back in January 2024 as a reverse engineering challenge/learning project and after posting some short videos on YouTube to show my friends the progress I've made, someone from another (now dead) IM revival project's Discord server found those videos and contacted me.

Turned out there is a fairly large community around old IM clients and some revival projects that already offered support for older Y!M versions (v5, v6), but with a very limited feature set (logging in, adding/removing friends and chatting).

Most if not all of these projects revolve around MSN/WLM, AIM and other IM clients that were more popular in the US. The Y!M support was typically an afterthought.

After some convincing from the guy that first contacted me on YouTube, I've decided to make this project public and with his help we've launched it in October 2024 as Buzzd Chat along with a Discord server where people could contact us and share their IDs and connect with each other.

I've started the project reversing Y!M V9 -- was the one I remembered using the most back in the day -- and because of that, the project got, for what it is and little to no publicity, a lot of attention. The subset of users interested in Y!M really wanted better support and newer versions.

The project is still under development, we've still got a long way to go till I'll be fully satisfied and ready to open source it.

Supported Y!M versions: - V9 only atm (focusing on implementing as many features as possible before exploring older or even newer versions)

Features (so far):

Authentication

Profile/Display pictures

Status management

Buddy management (add/delete/group/move buddies)

Buddy ignoring

Visibility management (invisible to all, invisible to some etc.)

Buddy list sharing (share your contacts)

Address book/Contacts management (add/remove/update contacts)

Photo sharing

File sharing

Audibles

Emoticons (all of them, even the ones introduced after V9)

Reliable messaging (basically acking that messages have been received)

I’ve been working on UHOP (Universal Hardware Optimization Platform) — an open-source framework that helps developers optimize GPU and accelerator workloads across architectures (CUDA, ROCm, OpenCL, etc.) without vendor lock-in.

It started as a personal frustration: I’d write something that ran great on CUDA, then have to rewrite or retune for ROCm or OpenCL. UHOP tries to make that portable — it detects your hardware, generates or benchmarks candidate kernels, and caches the best performer. It also supports AI-assisted kernel generation using OpenAI APIs and comes with a simple CLI for demos and benchmarking.

Right now, UHOP can:

Auto-detect hardware backends and pick optimal kernels

Run and benchmark fused ops like conv+ReLU

Cache and reuse tuned kernels

Generate kernels dynamically via codegen (CUDA/OpenCL/Python/Triton)

There’s still a lot in progress — better backend integration, distributed optimization, and a web dashboard for visualizing results. I’m sharing it early to get feedback from folks who’ve worked on compilers, GPU runtimes, and ML infra.

Repo: github.com/sevenloops/uhop

Demo: uhop.dev

Would love any thoughts on architecture, testing approaches, or potential contributions from ex-NVIDIA/ROCm engineers.

I built UHOP (Universal Hardware Optimization Platform) because I was tired of the invisible cage that GPU developers live in — everything we write seems to orbit around NVIDIA. CUDA is amazing, but it’s also a moat. Porting the same code to ROCm or OpenCL usually means starting over.

UHOP is an open-source attempt to break that lock-in by introducing a cross-vendor optimization layer. It detects your hardware (CUDA, ROCm, OpenCL, etc.), generates or benchmarks kernels, and caches the best performer. You can wrap your ops with a decorator, let UHOP choose or generate the kernel, and it just runs — wherever.

Features so far:

Hardware detection + backend selection

AI-assisted kernel generation (CUDA / OpenCL / Triton)

Fused op demos (conv2d+ReLU, matmul, etc.)

Kernel benchmarking and caching

CLI + early browser dashboard

There’s a long way to go — distributed tuning, compiler IR passes, better PyTorch/JAX hooks — but it’s open, hackable, and community-driven.

Repo: github.com/sevenloops/uhop

Demo: uhop.dev

Would love feedback from compiler engineers, GPU devs, or anyone who’s ever felt boxed in by vendor APIs.

I built this after hitting an infinite loop pagination trap on Newegg and wasting hours building a scraper that never had a chance.

caniscrape analyzes a URL and tells you: - What protections are active (WAF, CAPTCHA, rate limits, TLS fingerprinting, honeypots) - Difficulty score (0-10) - What tools/approach you'll need

Install: pip install caniscrape Usage: caniscrape https://example.com Web version: https://caniscrape.org

NOTE: Very hard sites like Amazon or YouTube will show as 0 because their bot protections are essentially invisible, I will fix this in a future update.

NOTE FOR WEB VERSION: The web version uses a cloud IP which is blacklisted by major sites. I will add proxy rotations to defeat this but for now, the website is only good for easy/medium sites.

Hi HN!

I built Streaky to solve a personal problem - I kept losing my GitHub streak on busy days. It monitors your contribution streak and sends notifications to Discord/Telegram before it breaks.

*What makes it interesting technically:*

1. *Distributed Cron Processing*: Used Cloudflare Service Bindings to bypass the 30-second CPU limit. Each user gets processed in an isolated Worker instance with its own CPU budget.

2. *Idempotent Queue System*: D1-based queue with atomic operations prevents duplicate processing when cron jobs overlap or retry.

3. *Zero-Knowledge Security*: GitHub tokens never stored (OAuth refresh flow), webhooks encrypted with AES-256-GCM, notifications sent via isolated Rust proxy.

4. *Rate Limit Solution*: Cloudflare Workers use shared IP pools which trigger rate limits from Discord/Telegram. Solved by routing notifications through a dedicated Rust server on Koyeb.

*Tech Stack:* - Frontend: Next.js 15, React 19, TypeScript - Backend: Cloudflare Workers + D1 (SQLite) - Infrastructure: Rust notification proxy - Auth: GitHub OAuth via NextAuth.js v5

*Live demo*: https://streakyy.vercel.app

The project is fully open-source under MIT license. Happy to answer any questions about the architecture or implementation!

We’re building Cloudtellix an AI FinOps assistant for AWS that gives recommendations with reasoning trails and integrates into existing workflows. Prototype + waitlist: [cloudtellix.com]. Would love technical feedback!

This beta includes several technical improvements aimed at performance, flexibility workflows:

Unlimited canvas: Previous canvas limits were removed. Earlier restrictions caused unnecessary re-renders and blocked operations. Now designers can place elements anywhere without overhead.

Optimized state management: Creating, updating, and deleting elements now triggers minimal state updates. Only affected parts of the state are recalculated, reducing processing and improving responsiveness.

Dedicated icon element: Icons now have their own element type. This decouples them from general elements, allowing integration of additional libraries, dynamic styling, and more flexible adaptation within the app.

Recursive in-memory export: Exports are built element by element recursively in memory. This ensures data consistency, avoids redundant recalculation, and scales efficiently to high-resolution outputs (up to 8K).

Mouse wheel zoom: Implemented a smooth zoom mechanism controlled via the mouse wheel for better navigation of complex layouts.

Multi-element property updates: When multiple elements are selected, only the modified property is updated for all elements. Full state recomputation is avoided, improving performance in large designs.

Auto-focus shortcut (Shift + 1): Elements are located by their x, y, width, and height, then auto-focused on the canvas. This prevents users from losing context while designing complex screens.

Tech stack: React, Next.js, MongoDB, Jotai, React Konva.

Feedback Requested: Looking for feedback on performance, UX, and data model structure.

Link: https://pixel-kit.vercel.app

Hey HN,

I've been building websites for 15 years. I'm a perfectionist - both visually and technically - and got tired of going back and forth between different tools, manual checks, and documentation for every audit.

Built GuardianScan to run everything in under 45 seconds: - Core Web Vitals (LCP, INP, CLS) - WCAG 2.2 compliance - Security headers and CSP policies - Modern framework patterns - SEO and schema markup

47 checks total.

These standards exist for a reason. Google prioritises fast sites. Accessible sites reach more people. Secure sites build trust. When you meet modern standards, you see it in your traffic and conversions.

Stack: Next.js 15, React 19, Supabase, Browserless.io for headless Chrome.

Won't catch everything - automated accessibility gets about 70% of issues. But it's what I use before every deploy now.

Landing page is up. Launching Nov 1. £24/month, no enterprise pricing BS.

Happy to answer questions.

Generate production-ready OpenAPI specs, frontend & backend code and documentation with AI

Hey HN, I vibe coded this in a day last time https://johnathannader.com/my-microphone-didnt-work/ and launched it; https://news.ycombinator.com/item?id=37503569, now that the coding para-dime has gotten better, decided to spend another day updating it, this is the result