2026年6月17日 的 Show HN

Hey HN, I built a website to watch live baseball games in an 8-bit broadcast. It takes live MLB data streams and converts them into near real-time pixel art gamecasts.

Been waiting to share this for when there’s actually a good slate of games happening since the site is pretty bare otherwise.

Here is today's schedule:

Mets @ Reds - 9:40am PDT https://ribbie.tv/watch/game/824503

Royals @ Nationals - 10:05am PDT https://ribbie.tv/watch/game/822721

Marlins @ Phillies - 10:05am PDT https://ribbie.tv/watch/game/823450

Tigers @ Astros - 11:10am PDT https://ribbie.tv/watch/game/824178

Padres @ Cardinals - 11:15am PDT https://ribbie.tv/watch/game/823044

..and another 14 games throughout the later day.

I'm still early on in this project, but I've tried to add little details with actual stadiums, day and night modes, between inning graphics and interstitials, live scoreboards, etc.

Would love any feedback and ideas. Thanks for checking it out!

Neural CAs model self-organizing pattern formation.

Now they can generate patterns at HD resolution in real-time, enabled by turning each CA cell into a Neural Field.

Try 3 demos: grow a pattern from a seed (and damage it, it heals), synthesize PBR textures that can regenerate, or create 3D textures like clouds.

Hey HN,

I'm Jay, co-creator of Mira. An open-source, self-hosted AI code reviewer where you BYOK (bring your own key).

Local models are getting really good.

Hosted frontier models are getting really expensive.

So, we built Mira. Mira has some snazzy features too:

- It's really fast at reviewing. Average is 77s compared with Greptile's 5 minutes. Your PRs aren't going into a queue on a cloud somewhere.

- Mira performs a blast radius and see what damage will be done with code change.

- Learns your codebase's patterns from the repo itself and enforces them without a config file.

- It's self-hosted, so your code never leaves your infrastructure. Bring your own model - OpenAI, Anthropic, or a local LLM behind your firewall.

We're really active, and looking for active contributors. Come join us in our discord!

https://github.com/miracodeai/mira

Thanks, Jay

Hi HN,

I spent the last few months reading the RISC‑V specification to build the lightest possible sandboxes. The idea behind a vpod is to quickly spin up a Linux sandbox from snapshots (Alpine by default) without any setup or subsystem required.

The trade-off for portability and security is raw CPU speed. So we don't expect it to match native workloads with Python or pip, for example.

More info is in the README https://github.com/capsulerun/vpod

Happy to answer any questions!

Hey HN,

There’s been a lot of interest recently in meta-harnesses, loops, and multi-agent orchestration. Obviously, there are already a lot of good tools: Conductor, cmux, the native Codex / Claude Code apps, etc.

For my own use cases, I’ve felt that the orchestration layer tends to feel overengineered. I mostly wanted a simple local harness (i.e Pi) for running and tracking CLI agents with the ability to hop in (via tmux). Relaymux is my opinionated attempt at that.

A few design principles:

- The frontend is just Telegram / iMessage / CLI. If I want more visibility, I hop into tmux.

- Subagents are normal interactive CLI agents running in tmux windows, usually with their own worktrees.

- The harness owns the tmux session, so each longer task becomes a named tab/window. Subagents report back to the orchestrator via CLI when they’re blocked or done. Then the orchestrator just messages me on Telegram / iMessage

- It works with any CLI agent that has an interactive terminal mode, so I don’t need special print-mode/non-interactive support. This means I don’t need to stress about the Agent SDK / claude -p billing limitations.

I love dosa, but have a hard time remembbering what all of the terms mean and how they fit together. "Mysore masala paper roast" is just a stack of independent choices: batter, texture, a smear, a filling, and the word dosa. So I built a little tool that treats every dosa name as a formula. You build an order and see what would arrive, or tap a real menu item and watch its formula light up.

I tried to (with Claude's help, obv.) cover naming conventions from Bengaluru, Chennai, and Hyderabad. And, full disclosure, I'm an outsider to this cuisine who just loves food, so there's every possibility I've got mistakes in there, and I'm certainly missing some great dosas. Feedback and corrections are very welcome!

Hi HN, for the past few months, we’ve been working on Ferrix AI (https://ferrix.ai/)

As AI agents speed up engineering, deciding what to build has become the bottleneck. Developers got faster because agents fit into their workflow: tech design, code review, testing. They still make the key decisions, while AI handles mechanical work around them.

At Ferrix, we are automating workflow for product managers to make product decisions efficiently. Agents handle the research synthesis, spec writing, and progress tracking, while PMs, designers, and engineers collaborate in a shared workflow.

Ferrix AI is live, and you can start free → https://app.ferrix.ai/

Hey HN, I built Reyn - which I like to describe as "granola but for everything". You're probably thinking another screen capture AI tool (which is true). Same as always, the biggest question that comes up is privacy, so I'll talk about that first

1. raw screen data is never stored in the cloud 2. user controlled filters are granular to the point that you're able to configure specific apps, windows, websites, or even keywords to be discarded immediately (once again never leaving your mac) and never captured down the pipeline

I personally built it because I find it useful and always had the problem of organizing my day (not note taking or task management), as well as sharing context on things that just happened to go undocumented throughout my day. As I was building it I decided to go even further and see if I could collect useful insights and find room for improvements in my day to day workflow.

This led to the current version of Reyn and its differentiating factor being the fact that it has a proactive layer. Most tools in this space are reactive - you ask, they retrieve. Reyn surfaces insights on its own and sends a daily recap of what you worked on, what's still open, and what deserves attention. The journal feature also lets you search across basically anything you've done on your Mac.

The proactive insights work by first having you configure what your ideal workday looks like — whether that's hours worked or the type of work being done. We have a few broad categories that tasks fall under, with more customization coming.

Current integrations:

Obsidian (available now, improvements in progress) Gmail, calendar, web search via a floating window with some agentic functionality Notion (coming soon) BYOK for LLM API requests (on the roadmap) ... and more

It's still early, but the journal and insights features are the strongest parts right now. Would love some feedback especially on the privacy model. My personal take - I think with enough safeguards in place, the data aggregated about your work is fully in your control. A lot of these data sources already store your data. If you're using Notion, Claude, or just browsing a website, that data is already being stored somewhere. Reyn is just aggregating it and putting it to work for you.

Happy to answer any questions about how it works

usereyn.com (public beta)

I've created an in-browser Python/Pandas/Git practice environment for my online learning platform and also for my corporate training classes. I'd be happy to discuss how I went about designing this, how I'm using it in my classes, and the architectural decisions I've made.

Most interesting, to me, is how much is running in the browser. Thanks to Svelte, Pyodide, isomorphic-git, LightningFS, and CodeMirror I'm able to provide a full environment for Python, Pandas, and Git.

I built much of this with Claude Code, and I'm happy to discuss how that went — what worked well and where I had to step in and make the calls myself.

I'm especially excited about the Git simulator: it shows the commit tree change as you run commands, plus an animated view of how files move between the working tree, the staging area, and HEAD.

The AI tutor, which uses Claude Haiku, was given my newsletters, classes, and exercises as inputs, along with a description of my pedagogical approach: instructors should give hints and feedback, but not reveal the answer.

When it comes to chess openings I tend to forget some lines in crucial situations mid game. To cure my problem I used a site called chessreps.com for drills and practice. After wanting to explore more lines I came across the paywall behind it, which left me with one option: To build the open-source version myself and finally its here: Tabia. Tabia at its core is a browser-local, no account, no server chess opening driller where nothing leaves your machine built with chess.js + Stockfish-in-WASM.

Hi HN!

I build this project to explore an idea I got in mind for a long time : Is transformer a suitable architecture for a chess bot? I built a small model (11M parameters) and trained it on human games (Elite Lichess DB).

Model alone is performing around 1500 elo, but I built an harness using Monte Carlos Tree Search (MCTS) using my model heuristics to improve the model to ~2100 elo (evaluated against stockfish).

If you want to try it, it is available as a Lichess bot : https://lichess.org/@/ChessTransformerBot

I'm looking to evaluate this model against human players so challenge, I would be grateful if you try it!

The project is open source, don't hesitate to star the repos if you like the project.

For me, the main key learning is that machine learning is an important part of the project, but it was the harness design that makes the system works with a nice performance regarding the small model size.

Hello HN! I made an iPad app with my seven-year-old daughter to make learning spelling fun.

https://spellabee.com/

We play Spelling Bee type games in our car rides, and she wanted to learn more words that way. So we built a simple app that teaches 10 words at a time, and lets the kids practice and master these 10 words.

The full word list in the app is static, and it gets progressively harder as the kid goes through the levels. There are no AI features in the app. I do not collect emails inside the app or have third party trackers. Based on feedback (reviews) and aggregate usage data I plan on updating the app with new word sets.

Although the app does not have any AI features I used AI to build the app itself. I used Claude to code the app using Flutter, do etymology research, and understand what alternative apps that are in the App Store. While the LLMs were good at providing a lot of information, I had to synthesize it and play a strong Product Manager role to drive the vision and keep the app simple. My daughter provided a lot of feedback and helped simplify the app and refine the UX. The "Bee Stage" design was inspired by her drawing.

Without AI tools, it would have been almost impossible for me to build and launch this app. But it still required a lot of decision-making and prioritization to get the product out of the door. I strongly believe that while AI is a powerful tool, human taste is the differentiator in well made products.

If you have a kid in K-5 who is interested in spelling bee type games, give it a try and I would love to hear any feedback you may have as a parent.

App store: https://apps.apple.com/app/apple-store/id6768881287?pt=12867...

Hey all, I've built a web app for practicing spoken conversations in a language you're learning. It was driven by my own needs when I learn Korean. The main functionality includes two modes: 1. Chat with AI — you pick a situation (ordering at a café, taking a taxi, checking into a hotel) and the AI plays the other role entirely in your target language. It stays in character, adapts to what you say, and you can fail as many times as you want with nobody watching.

2. Chat with Human — for real, in-person conversations with native speakers (think traveling abroad and talking to a local). It translates the conversation in real time so both sides can actually communicate, and saves the full transcript so you can review it afterward and learn from what was said.

I'm a solo dev. The web app is live today and mobile is still being built. The app is still at early stage, I'm adding more features to it. I'd love some feedbacks, especially on what features you think are missing and what you would expect from a language learning app.

Here is a demo page where you can play with it without signing up: https://www.chattylingo.com/demo

Hi HN, I wanted to share a project I've been working on recently.

I built WaylandClientKit mostly for fun and curiosity, and to explore a gap I felt existed in the Swift world which is native Wayland/Linux desktop foundations.

Not as flashy as a UI toolkit you'd build on top - more of a lower level substrate for windows, input, text input, data transfer, cursors, and preview graphics APIs. Its still early, but I'd love any feedback.

"GRID" (Gaps, Risks, Inconsistencies, Discrepancies) is a small framework that I developed over time for my spec-driven development using coding agents.