Show HN for July 11, 2025

Hey HN, Kyle here, one of the co-founders of OpenPipe.

Reinforcement learning is one of the best techniques for making agents more reliable, and has been widely adopted by frontier labs. However, adoption in the outside community has been slow because it's so hard to implement.

One of the biggest challenges when adapting RL to a new task is the need for a task-specific "reward function" (way of measuring success). This is often difficult to define, and requires either high-quality labeled data and/or significant domain expertise to generate.

RULER is a drop-in reward function that works across different tasks without any of that complexity.

It works by showing N trajectories to an LLM judge and asking it to rank them relative to each other. This sidesteps the calibration issues that plague most LLM-as-judge approaches. Combined with GRPO (which only cares about relative scores within groups), it just works (surprisingly well!).

We have a full writeup on the blog, including results on 4 production tasks. On all 4 tasks, small Qwen 2.5 models trained with RULER+GRPO beat the best prompted frontier model, despite being significantly smaller and cheaper to run. Surprisingly, they even beat models trained with hand-crafted reward functions on 3/4 tasks! https://openpipe.ai/blog/ruler

Repo: https://github.com/OpenPipe/ART

Hey HN. we’re Barnaby and Tom from Director (https://director.run). A fully open source, local first MCP gateway that allows you to connect Claude, Cursor or VSCode to any MCP server in 30 seconds.

Demo: https://www.youtube.com/watch?v=Fj4mGLp-aSk

Website: https://director.run

Repository: https://github.com/director-run/director

MCP is a promising technology, but it's still early and there are lot's of (well known) problems:

- Configuration: Hard to setup, requiring writing JSON for each new client <> server connection.

- Observability: No easy way to inspect (and modify) traffic flowing between clients and servers.

- Security: MCP is often vulnerable to remote code execution attacks and prompt injection attacks.

- Context Window Management: It's very easy to load too many tools into the context window at which point the LLM can get confused.

We're looking to solve these problems by building director. We're starting with a simple local experience, but intend to add features like oAuth very soon that will allow you to run this in the cloud (or in a secure container).

Looking forward to hearing any feedback or ideas you have!

Hi Hacker News,

I'm excited to share a project I've been passionately working on in my spare time: AI Movie Finder (https://www.aimoviefinder.com).

The idea was born out of a simple, yet frequent, frustration: having a movie on the tip of my tongue but being unable to recall the title. All I could remember were fragments – a specific scene, a snippet of dialogue, the general plot, or even just the mood it evoked. Traditional search engines often fell short with these kinds of abstract queries.

So, I decided to build a solution. AI Movie Finder uses a natural language processing model to understand these descriptive and sometimes vague queries. You can type in things like:

"that movie where a guy keeps reliving the same day"

"a sci-fi film with a blue alien opera singer"

"a feel-good movie about a band in the 80s"

The goal is to make movie discovery more intuitive and human-like. Instead of just searching by actors or exact titles, you can search by memory and feeling.

The backend is built with Python and utilizes a fine-tuned sentence transformer model to create vector embeddings for a large movie database. The frontend is a clean and simple interface built with vanilla JavaScript to keep it fast and accessible.

This is still very much a work in progress, and the database is continuously growing. I'm actively working on improving the model's accuracy and expanding the search capabilities.

I would love to get your feedback. Please give it a try and let me know what you think. I'm particularly interested in:

How well does it work for your queries?

Are there any features you think would be a great addition?

Any suggestions on how to improve the model or the user experience?

Thanks for checking it out! I'll be here to answer any questions.

Building a free web toolbox for hardware engineers: harness cable diagram generator, connector pinout tools (M.2, JTAG...), microcontroller serial monitor, and various other small utilities.

No sign-up required. Designed from my own needs. Feedback and feature suggestions are welcome.

Hey HN!

I've been building CSV editors for 15 years. This is my third rewrite, and I think I finally got it right.

Key features: - Excel-like operation for intuitive use - Equipped with basic to advanced tools necessary for handling CSV - Supports various formats and character encodings. Can handle CSVs with different column counts - Cross-platform (Windows/Mac, Linux coming soon) - Opens 100MB files in 1.6 seconds (12x faster than Excel)

Tech stack: Tauri (Rust + React/TypeScript/TailwindCSS)

Would love feedback from anyone dealing with large datasets daily!

Website: https://smoothcsv.com/ Blog: https://dev.to/kohii/the-technology-behind-smoothcsv-the-ult...

I wanted a simple way to manage my monthly budget, and I got tired of trying to bend Google Sheets to do what I needed.

So I built a terminal-based personal budget planning application built with Python and Textual.

Your data is stored locally in a JSON file, keeping your financial information private and accessible offline.

Do you find this useful?

With Gemini Flash so fast, I wondered what it would be like for an LLM to generate web pages and images on-demand as the URLs are requested. It's been a couple of weeks now since release and there are a ton of cool examples people have created at https://ginprov.com/. I have about half of my Gemini credits left (it's not too costly) but if it runs out, it's very easy to self-host with your own Gemini key.

Some examples:

https://ginprov.com/dachsunds/

https://ginprov.com/cool-cars/

https://ginprov.com/coffee-in-a-golf-ball/

https://ginprov.com/what-is-it/

Hi!

There may be times when you want to generate multiple barcodes and scan them (even if there is not now, there may be such a scene in the future).

In that case, this site will be useful!

https://ddddddo.github.io/barcode/

This site can generate QR codes from URLs and multiple barcodes! The barcode generation process is Go! (For now, it is only displayed in Japanese.)

The features and functions of this site are detailed below.

- Code generation process is performed in Wasm - The code generation process is completed locally, without communication with external parties. - Code generation process uses github.com/boombuler/barcode

- Multiple barcodes can be generated on a single page - Enter a URL in the URL form to generate a QR code for that URL. - Enter the Basic Authentication credentials for the target URL in the User/Password form, and the credentials will be embedded in the generated QR code. - Click the “Add Barcode” button to generate multiple barcodes. - QR Code for URL can be hidden. - If the QR code for the generated URL contains Basic Authentication information, the page cannot be shared in a screenshot, so you can hide the QR code by pressing the dedicated button. - Enlargement and blurring on mouse-over - When there are multiple barcodes, you may scan another barcode when scanning. In such cases, mouse-over the target barcode to enlarge and blur the other barcodes, making it easier to scan. - Share page content by URL - The query parameters reflect the URL entered in the form and the string from which each barcode was generated. Use it to restore this page - However, User / Password is not reflected in the query parameter because it is not allowed to be shared.

The repository for this site is https://github.com/ddddddO/barcode

I built nkv because I wanted a simple way to pass simple values between different devices without maintaining a central server. Nostr provides a backbone with "good enough" decentralization. Enjoy!

Hi HN,

Over the last two months, I’ve developed a model where computation is realized as symbolic states embedded directly on geometric manifolds—specifically, helices with torsion, chirality, and combinatorial structure.

In a nutshell - using geometry you can encode computational logic into the rotation group of SO2 (the circle) via the torsion of the helix. Helices are always oriented shapes, so their unit vector/axis of symmetry is used to determine the logic of the system. This deterministic approach stands in contrast to quantum computing where you use SU2. Likewise, the gluing maps between helical manifolds can create what I call "computational geometric bordisms".

This approach deliberately sidesteps the classical “coordinate-free” dogma in math, instead offering a deterministic geometric substrate for computation that naturally interpolates between classical, quantum, and topological computation.

The results are honestly pretty strange: potentially foundational, deeply cross-disciplinary, and—full disclosure—I’m not a professional mathematician. My intent here isn’t to claim a “Theory of Everything,” but rather to provide a teaching tool for understanding computation through geometry.

However, in working through the details, this model seems to (a) threaten to “destroy” one of the Clay Millennium Prize Problems, and (b) force a substantial extension of Grothendieck’s Homotopy Hypothesis. That wasn’t my original goal, but it’s hard to ignore the consequences.

I’m really looking for a sanity check, constructive critique, or any pointers to researchers working along similar lines. Feedback—critical or otherwise—is genuinely appreciated.

Repo here: https://github.com/LambdaLord/Twin-Helix-Geometric-Computati...

Abstract below:

Geometric Computability: Twin Helices and the Algebra of Flow

In this work, I construct a manifold comprising of twin helices that explicitly validates the existence of computational geometric bordisms as anticipated by the project of the cobordism hypothesis. I investigate its geometric, topological, algebraic and computational properties. This work suggests the emergence of a new research direction that I call “Geometric Computability Theory”, situated at the crossroads of geometry, topology, and computability theory. I prove that encoding of arbitrary size finite state machines is possible using these “Geometrically Computable Manifolds” leading to the emergence of Turing Complete phenomena. Further constructions are made on the nature of quantum-like and n-ary/fuzzy computing behavior leading to the formation of a general metatheorem for the behavior of computational logic to explain the origin of Rice’s Theorem. I describe their behavior in terms of Kleene’s theory of partial recursive functions and use vector spaces and ordinary differential equations to model fluidic computation. This perspective opens numerous avenues for future investigation, potentially reshaping our understanding of computation in continuous spaces including applications to the infamous Yang-Mills existence and mass gap. As of writing, algebraic systems to formalize the behavior these geometrically computable manifolds appears to be absent. Future work will involve the development of formal algebraic frameworks, rigorous computational models, and experimental simulation of system behaviors.

Hey HN,

I'm excited to share Amebo, a new open-source HTTP Event Broadcasting library I've been working on. It's designed to act as an asynchronous communication engine that helps decouple your applications from traditional, often complex, messaging systems like PubSub, Kafka, RabbitMQ, or SQS.

The core idea behind Amebo is to provide a simple, HTTP-first API for registering event schemas, submitting message payloads, and broadcasting those events to various destinations. This makes building event-driven architectures or microservices much more straightforward, without the overhead of managing dedicated message brokers for every interaction.

Key highlights:

High Performance: Built for scale with sub-10ms latencies, efficient batching, and connection pooling. It's been battle-tested with millions of requests.

Simple Integration: JSON Schema validation for events, flexible backend support (PostgreSQL, SQLite, Redis), and a clean RESTful design.

Enterprise-Ready: Features like clustering for high availability, built-in authentication/authorization, and observability hooks are baked in.

Flexible Event Delivery: Supports various "Event Engines" for broadcasting, including PubSub, Kafka, RabbitMQ, SQS, and direct Webhook delivery with retries.

Amebo is ideal for scenarios where you need robust, asynchronous communication between services but want to avoid deep integrations with specific messaging queues. Think event sourcing, lightweight service-to-service communication, or reliable webhook delivery.

We've focused on making it easy to deploy (Docker-ready) and simple to integrate into existing Python ecosystems.

You can find the documentation and source code here:

Documentation: https://rayattack.github.io/amebo/

GitHub Repository: [Your GitHub Repo Link Here - e.g., https://github.com/RayAttack/amebo]

I'd love to hear your thoughts, feedback, or any questions you might have. Thanks for checking it out!

Hi Hacker Family,

I’d love to share a project I’ve been building independently: AI Email Writer – WriteMail.app

The inspiration came from a personal experience. I was applying for a part-time role at a company overseas, but I wasn’t confident about how to write a proper email. Sure, there are many AI tools out there, but I found it frustrating to constantly craft the right prompts just to get a decent draft.

That got me thinking — maybe others face the same challenge.

So I created WriteMail.app – an AI-powered email writing lab focused on helping you write better emails, not just faster ones.

Unlike typical “email generators,” WriteMail is a guided and interactive platform that blends writing purpose, tone control, and message structure — so your emails sound more professional, effective, and human.

What you can do with WriteMail.app:

1. Quickly write business emails, requests, rejections, thank-you notes, and more

2. Define your intent, and let the AI choose the right tone and phrasing

3. Instantly adjust the tone – make it more polite, assertive, concise, etc.

4. Need to reply to a message? Paste it in — and get a thoughtful response

5. Multilingual support is built-in – write or reply in the language you prefer

We’re currently powered by GPT-4, and I’m actively optimizing the system to fine-tune prompts and outputs for even more accurate and polished results.

It’s still an early-stage project, but I’d be thrilled if you give it a try and share your feedback. I’m building this to make thoughtful email writing more accessible for everyone.

Really appreciate your time, and I’m excited to hear any feedback or suggestions you may have.

Independent images collection from the whole internet. Daily updates