Show HN за 17 сентября 2025 г.

Hi HN,

I’m the creator of Math2Tex. I was a PhD student, I spend a huge amount of my time working with LaTeX, especially when dealing with lecture notes, academic papers, and homework. I built *Math2Tex*, a lightweight tool that converts handwritten or printed academic content — especially math formulas — into LaTeX or text

The Problem:

I've always found it incredibly tedious to manually type out mathematical formulas, especially complex, multi-line equations from my handwritten notes or from a textbook. It's slow, boring, and I always make syntax errors. I tried some existing tools, but they often struggled with my handwriting or couldn't handle mixed content (text and formulas together).

The Solution:

So, I built Math2Tex to solve my own problem. It’s a straightforward, single-page web app: you upload an image (a photo of your notebook, a screenshot of a PDF, etc.), and it converts the academic content into clean LaTeX code or plain text. You get a real-time preview and can copy the result with one click. My goal was to make the workflow as fast as possible: Snap. Convert. Done.

You can try it here: [https://math2tex.com](https://math2tex.com/)

How is it different from general AI tools like GPT, Claude, etc?

This is a fair question. While large models can handle this, they are often slow for such a specific task. I wanted something faster and more specialized. Math2Tex uses a lightweight model fine-tuned specifically for academic content recognition.

In short, think of it as a specialized scalpel versus a Swiss Army knife. For this particular job, it's generally 3-5x faster and, in my experience, more reliable for complex notations.

Tech Stack:

The core OCR engine is a custom-trained model based on a transformer architecture, fine-tuned on a large dataset of both printed and handwritten academic material. It's all deployed on Vercel.

*It's free to use.* This is still an early version, and I'm sure there are plenty of bugs and areas for improvement. The recognition might not be perfect, especially with very messy handwriting or some obscure symbols.

I would be incredibly grateful for your feedback. Whether you’re a student, researcher, or someone who’s fought with LaTeX input. Feedback on both the tool and the approach would be really helpful.

Thanks!

So, after years of abandoning Rust after the hello world stage, I finally decided to do something substantial. It started with simple line rendering, but I liked how it was progressing so I figured I could make a reasonably complete PSX style renderer and a game with it.

My only dependency is SDL2; I treat it as my "platform", so it handles windowing, input and audio. This means my Cargo.toml is as simple as:

[dependencies.sdl2] version = "0.35" default-features = false features = ["mixer"]

this pulls around 6-7 other dependencies.

I am doing actual true color 3D rendering (with Z buffer, transforming, lighting and rasterizing each triangle and so on, no special techniques or raycasting), the framebuffer is 320x180 (widescreen 320x240). SDL handles the hardware-accelerated final scaling to the display resolution (if available, for example in VMs it's sometimes not so it's pure software). I do my own physics, quaternion/matrix/vector math, TGA and OBJ loading.

Performance: I have not spent a lot of time on this really, but I am kind of satisfied: FPS ranges from [200-500] on a 2011 i5 Thinkpad to [70-80] on a 2005 Pentium laptop (this could barely run rustc...I had to jump through some hoops to make it work on 32 bit Linux), to [40-50] on a RaspberryPi 3B+. I don't have more modern hardware to test.

All of this is single threaded, no SIMD, no inline asm. Also, implementing interlaced rendering provided a +50% perf boost (and a nice effect).

The Pentium laptop has an ATI (yes) chip which is, maybe not surprisingly, supported perfectly by SDL.

Regarding Rust: I've barely touched the language. I am using it more as a "C with vec!s, borrow checker, pattern matching, error propagation, and traits". I love the syntax of the subset that I use; it's crystal clear, readable, ergonomic. Things like matches/ifs returning values are extremely useful for concise and productive code. However, pro/idiomatic code that I see around, looks unreadable to me. I've written all of the code from scratch on my own terms, so this was not a problem, but still... In any case, the ecosystem and tooling are amazing. All in all, an amazing development experience. I am a bit afraid to switch back to C++ for my next project.

Also, rustup/cargo made things a walk in the park while creating a deployment script that automates the whole process: after a commit, it scans source files for used assets and packages only those, copies dependencies (DLLs for Win), sets up build dependencies depending on the target, builds all 3 targets (Win10_64, Linux32, Linux64), bundles everything into separate zips and uploads them to my local server. I am doing this from a 64bit Lubuntu 18.04 virtual machine.

You can try the game and read all info about it on the linked itch.io page: https://totenarctanz.itch.io/a-scavenging-trip

All assets (audio/images/fonts) where also made by me for this project (you could guess from the low quality).

Development tools: Geany (on Linux), notepad++ (on Windows), both vanilla with no plugins, Blender, Gimp, REAPER.

Would love any feedback

For speech-to-text, large-language-model inference and text-to-speech I created three wrapper libraries in C/C++ (using Whisper.cpp, Llama.cpp and Piper).

Follow the URL to see an example that shows how to use these libraries for a speech-to-text, LLM inference, text-to-speech pipeline.

Windows and Linux are supported.

Hey all! npm-daycare is a simple NPM proxy built on Verdaccio which filters all packages that:

- are younger than 48h (it will just provide the old version instead)

- have fewer than 5,000 weekly downloads

This is in response to the recent supply chain attacks that shattered the JavaScript ecosystem [1]. It's likely not a problem that will go away any time soon, so we figured we'd build something to protect against it.

Doing this on the proxy layer means it will work across the entire system, as proxies are set globally. In the future, we could also add more filters to the proxy.

To get started, just run the Docker container:

    docker run -d --rm --name npm-daycare -p 4873:4873 bgodil/npm-daycare

    npm set registry http://localhost:4873/
    pnpm config set registry http://localhost:4873/
    yarn config set registry http://localhost:4873/
    bun config set registry http://localhost:4873/

    npm view @types/node  # has recent updates
    npm view pgmock  # has <5,000 weekly downloads

Downside: npm-daycare won't show packages that are younger than 48h on its default config, so be aware of that when you try to update your packages to patch a zero-day exploit.

You probably also shouldn't rely on this as your only line of defense. Curious to hear what you think!

[1] https://news.ycombinator.com/item?id=45260741

An offline first audio deck station

Does need online access but can play offline.

HTML5 needed.

Load local files, up to 2 GB audio.

Smooth transition between tracks.

EQ.

Compressor, pitch and speed controls.

Uses tone.js

I've been working on OrderlyID, a new identifier format for distributed systems.

It's like UUID/ULID/TypeID, but with a few twists:

- Typed: every ID has a human-readable prefix (order_xxx, user_xxx).

- K-sortable: lexicographic order ≈ creation time.

- Structured fields: 160-bit body includes time, tenant, shard, sequence, random.

- Checksums: optional 4-char integrity check to catch copy/paste errors.

- Privacy flag: can bucket timestamps for public-facing IDs.

Format:

<prefix>_<payload>[-<checksum>]

Example:

order_00myngy59c0003000dfk59mg3e36j3rr-9xgg

There's a draft spec with conformance tests:

https://github.com/kpiljoong/orderlyid/blob/main/spec/0001-s...

Go reference implementation and CLI:

https://github.com/kpiljoong/orderlyid

Compared to TypeID, OrderlyID adds larger bit size (160 vs 128), tenant/shard/sequence fields, optional checksum, and a privacy bucket flag.

Status: Draft v0.1 — stable enough for experimentation. Feedback and contributions very welcome.

Repo: https://github.com/kpiljoong/orderlyid

Hi HN,

I’m excited to share HuMo AI, an AI-driven tool that helps creators easily produce realistic, human-centric videos. HuMo AI supports text, image, and audio inputs, turning simple ideas into fully customized, lifelike content.

Key Features:

Multi-Input Support: Combine text, images, and audio to generate videos.

Realistic Results: Lifelike videos with perfect synchronization.

Perfect for Storytelling: Ideal for immersive experiences, education, and character creation.

Full Customization: Tailor every element, from appearance to actions.

HuMo AI uses an advanced AI reasoning engine, making it highly versatile for various creative tasks. Whether for gaming, education, or marketing, it offers a new level of freedom and control for creators.

The main challenge was integrating different input types to maintain synchronization and consistency. We’re continuously refining this, and we’d love your feedback.

Hi Hacker News ! Today I have new product waitlist that will launch soon next week.

Workser is Coding AI Agent API for Developers to integrate into your product to deliver vibe coding for your business use cases.

You can custom branding with prebuilt UI to make it your own vibe coding ai agent.

Started with Custom web app by ai agent. But soon with more types of software development.

Tell me how you think or join waitlist when we launch next week.

Hi all!

A few things of note -- I want this to remain a freemium service, where a reasonable amount of scans / capabilities are free per month etc. This is a passion project for helping people get access to security tooling that don't have the time/budget/etc to manage their own infrastructure. At the same time, I am paying a lot out of pocket to make this work, so I have added a credit system for users to support the cause at large.

I don't have much $$$ for bounties, but I am willing to pay what I can. If you find a security issue on the site, please report it via the contact info on https://meldsecurity.com/security.txt

I've got a roadmap of security tools to add to the toolbox, such as Dependency Scanners (E.g Trivy), Log Analyzers, secrets scanners, etc. If there's something you'd like added to the toolbox, let me know!

Also if you need more credits after the sign up bonus, just open a support case and ask for it and I will add some to your account! (obviously purchasing a credits package is greatly appreciated ;))

Anyways, any feedback would be greatly appreciated!

I often find myself trying to solve a geometry problem where the constraints are really simple to understand, but solving it algebraic is really hard and tedious.

I built this whole thing from scratch with Claude Code. It's my first time trying it and I literally did not write a single line of code...

That said, it still would be hard build this as a novice. I had to guide things along the happy path, but it saved me a ton of time!

The code is open source! Let me know if you run into any issues.

Hi Everyone!

Excited to introduce ya’ll to Navi!

Navi is a hackable smart speaker built for developers. All interfacing is done entirely through voice w/ Navi, who acts as an agent to complete any long term goal you care about, or simply to chat / be a SOTA research assistant.

Navi is completely customizable, you can add your own custom voice, change it’s personality, and hook in any custom tools you care about so as best to fit your needs (or pick from a dev marketplace). It also has a pretty killer memory system that consistently updates it’s world view to better help you achieve the goals you care about.

Would love any feedback or questions! I’m still super early on in the process & would love to hear about any of your thoughts :)

Give a child advice, clap back on the bully, and gobble a harmonious meet-cute in this game where you must compel the AI judge that what you said is the right thing to say

I built Vatify, a lightweight API to take the pain out of EU VAT. With just three endpoints you can validate VAT IDs, fetch up-to-date rates, and calculate the correct VAT (including reverse charge logic).

I hacked an AI background checker. After a week or two of nights & weekends with Cursor, React, Typescript, Supabase, Vercel, etc... the site is live.

Preview reports are redacted and full reports read a bit like Wikipedia pages for anyone.

The reports are quite fun. Try it out! If you or someone you know are interested in lots of reports, please feel free to reach out with a note about the use-case.

Cheers

Hi HN,

This is Isaac, co-founder of Leapcell (https://leapcell.io/).

I used to shut down small hobby projects because cloud costs and maintenance were too high. They would just sit on GitHub, unused.

Leapcell lets you keep your ideas online without worrying about cost. You can deploy *up to 20 projects for free* using our serverless container architecture, so multiple projects can run simultaneously.

Switch between modes as your project grows: - *Serverless*: fast cold start (<250ms), resources scale with traffic, ideal for quickly validating ideas. - *Dedicated machines*: stable, predictable costs for more mature projects, to avoid unexpected serverless bills.

This way, you can deploy all your code online, and Leapcell helps you find the most cost-effective setup at any stage.

Leapcell also provides built-in PostgreSQL, Redis, logging, async tasks, and web analytics, so your projects just work.