Ask HN: What are you working on? (May 2025)

I've been working on a drawing/animation library/language based on patterns and abstractions.

On one hand the idea seems so simple and intuitive (Define patterns (like 3 red blocks to the right), combine patterns ( 5 up * 3 red right), use patterns inside patterns (each block is a square), but implentation wise I keep running into so many intracies and I want it to be perfect so it's been kind of tough and slow.

I recently quit my salary job after 16 years and am consulting in nuclear engineering now. I have a few passion projects that I'm working on (between the somewhat substantial consulting work that came out of the woodwork):

- Nuclear Reactor Starter Kit --- an open source set of procedures, processes, templates, and maybe even some IT advice that should help newcomers start companies with nuclear quality assurance programs easily and quickly while also making a new format in which nuclear companies can share lessons learned in efficiency.

- Reactor Database --- similar to the iaeas PRIS but focused on reactor development rather than power reactors. Will include nuclear startup company tracking with details gleaned from statements and maybe extrapolated where necessary from simple simulations. Will include things like fuel cost and licensing progress. This way people can more easily separate vaporware from real nuclear, and keep track of promises vs delivery.

This is cool - how does IP geolocation work? How do you know that xyz IP is at this particular spot?

Edit: I see you are using MaxMind database - do you add some sort of additional analytics or overlay on top of that?

I've been collecting and digitizing vintage print advertisements and publishing them (https://adretro.com).

I have tens of thousands of ads in the collection and it would take me many lifetimes to complete, but I've been using AI to extract and catalog the meta data. I can get through about 100 ads/day this way.

https://tritium.legal

Tritium is an IDE for corporate lawyers. Draft Word docs, review PDFs, redline all in a single application. It's written in Rust using a modified version of egui. Immediate mode has some interesting tradeoffs that I'd love to discuss on here. Also the web/desktop dichotomy presents a lot of interesting opportunities and challenges where data governance is concerned. I'd love your thoughts or to share mine!

https://www.moviemixup.com

A wordle-like game based on a road trip game my friends and I used to play. It serves you up a mashup of two different movie plots, and you have to guess the combined movie title. There's always some sort of shared word or wordplay between the two movie titles.

An example from the tutorial: the day after tomorrow never dies.

On https://github.com/anttiharju/vmatch as a hobby. It's starting to get to a workable state, I'm using it to manage Go and golangci-lint for the project itself. It even works with the Go vs code plugin.

I think many version managers make things unnecessarily difficult, especially if one hops from one repo to another. vmatch automatically uses and installs the right versions.

I've been working on a calibration website / app.

Along the lines of predictionbook, metaculus - something that helps you be "well calibrated", but more playful/fun than metaculus.

It doesn't have a lot of upside - predictionbook actually went offline due to lack of interest. But it was a good excuse to try out some vibe coding, and learn react native (I've mostly been a backend programmer).

In an attempt to make it more engaging and fun, I decided to have it focus on sports picks. Also partly because calibration graphs need to have a lot of predictions to yield any reliable information about your calibration.

I got it up in time for March Madness and about 25 of my friends joined and it was a good time. I nagged and reminded them a lot about about 15-20 of them predicted all 63 games, by picking the winner of each match and what their percentage confidence was. I had a leaderboard and live-blogged and gave silly awards.

I later added support for multiple "tournaments" and currently have tournaments going for NBA Playoffs and NHL Playoffs, but interest is waning. Of my friends, only 2-3 others are still regularly predicting.

Maybe it'll be more fun for the NFL season but I might also let it go a bit dormant.

Biggest challenge is that there isn't really a bulletproof way to rank people if people only predict some games in a tournament. I've tried all sorts of things, minimum # of games, bayesian kernel smoothing, but it's ultimately arbitrary when choosing how to penalize someone for not participating.

If I were to continue I'd be looking at things like automatically integrating with sports apis and odds/bookmaking apis, allowing users to create their own tournaments, etc. But ultimately, the UX of the site isn't much more than making a prediction, and then checking back later when the game is over to see your score. Not much more reason to hang around on the site than that.

I was building an intentionally annoying app against doomscrolling [1]. Being technical, I tried to focus on product, marketing and more instead of the implementation. But I still didn't ship quick enough. It's so hard. Only after a few months did I start with marketing, and it hit me like a wall.

So I'm giving a try to a project which started with marketing. No implementation, just a TikTok to see if people like it. And holy crap, we got 75k views!

The new idea [2] is easier to explain (1 pushup = 1 minute of scrolling) and already has a community. Plus, not working alone helps me focus on what I'm good at: programming. I don't regret learning about other areas but doing marketing for a living is not my thing.

I'm not getting rid of SpeedBump, though. It's a fun side project and it does help people :)

[1] https://speedbumpapp.com

[2] https://pushscroll.com

Vibe coding a few apps I always felt humanity deserves (a bit exaggerated but kind of not :) )

- https://padsnap.app/ : PadSnap is a simple web app that adds customizable padding to your images so they fit Instagram’s/custom dimensions — no cropping, no quality loss. All on browser, no server uploads. Also no ads or login.

- https://shiryakhat.net/ : redid my podcasts website last week: Shir Ya Khat podcast, which translates to "Head or Tails" in Farsi, began its non-profit journey in 2016 with a mission to make blockchain and cryptocurrency technical knowledge accessible to Farsi speakers worldwide.

- life timetime visualizer, still WIP, feedback welcome: https://shayanb.github.io/timeline/

I’ve been experimenting with data formats like Parquet and Iceberg, and recently came across Lance. I wanted to try building something around it.

So I put together a simple Digital Asset Manager (DAM) where:

* Images are uploaded and vectorized using CLIP * Vectors are stored in Lance format directly on Cloudflare R2 * Search is done via Lance, comparing natural language queries to image vectors * The whole thing runs on Fly.io across three small FastAPI apps (upload, search, frontend)

No Postgres or Mongo. No AI, Just object storage and files.

You can try it here:

* https://metabare.com/

Or see the code here:

* https://github.com/gordonmurray/metabare.com

Would love feedback or ideas on where to take it next — I’m planning to add image tracking and store that usage data in Parquet or Iceberg on R2 as well.

Just released a “Loudness Contour” audio plugin. Let’s you apply various equal-loudness contours like Fletcher-Munson, ISO-226, LUFS style K-weighting, etc.

Fits into my “loudness series” suite of tools.

Have 3 more in development and then it’ll be on to the next series.

https://apu.software/contour/

Inspired by MathAcademy, I'm developing:

1) a note-taking workflow in Obsidian (you take bite-sized notes about a topic, then connect "prerequisite" notes in Obsidian's canvas editor)

2) a tool that uploads each note and graph data to a database

3) a webapp that presents those notes algorithmically using spaced repetition. This enables you to allow others to "traverse" your note graph in a guided and self-paced manner.

You can add "challenge presets" to each note so that your mastery of each piece of knowledge can be tested with simple flashcards, multiple choice, free response, or some visual/actionable task to force active recall. An algorithm uses your success rate and spaced repetition data to introduce & drill more advanced notes into your long term memory.

Here's some more reading I was inspired by:

https://www.mathacademy.com/pedagogy

https://www.justinmath.com/individualized-spaced-repetition-...

Even if there are a lot of imperfections and flaws about this project (like the sheer difficulty of curating a good knowledge graph to begin with), I'm hoping to make my note-taking in Obsidian more structured and thorough, replace my Anki routine, and make any of my notes into an automated + algorithmic course. If someone has another similar project (combining note-taking with hierarchal, topological knowledge graphs with spaced repetition and testing all in one platform) I would love to hear more about your approaches. Quick shoutout to one person I've seen who is doing something similar: https://x.com/JeffreyBiles/status/1926639544666816774

As far as what I'm focusing on this weekend:

1. Right now, working on standing up an MCP server in Java. Not using the Spring Boot support at the moment, but rather setting up embedded Tomcat and doing it the more "low level" way just for didactic purposes. I'm sure I'll use Spring Boot once I get deeper into all of this.

2. Plowing through the "AI Agents in Action" book. I'm just wrapping up the section on AutoGen and about to move into crew.ai stuff.

3. Reading a book on Software Product Line Engineering.

4. I have an older project that's Grails based that I let linger without any attention for a really long time. I'm working on updating it to run on the latest Grails and Java versions and also writing some automated smoke tests.

Free Resume Builder

When I was looking for a job last summer, I got frustrated with the current resume builders on the market and decided to build one exactly how I wanted to use it.

- No signup, no login, and no payment.

- Suggest a professional summary (with highlighting) to match a job description [0].

- Preview as you go.

- ATS friendly templates.

- Find relevant jobs for my resume.

[0] Recruiters skim through resumes, and highlighting the keywords they look for has always helped me to get their attention, so I decided to implement this feature using AI.

https://resumeyay.com

Unifi Video was replaced by Unifi Protect some time in 2020. I wasn't sure how to self-host Protect, so I never migrated to it. I've recently reached a situation where some phones can no longer install the Unifi Video app. These phones are now relegated to using the rough-on-mobile UI. The Unifi Video web UI has also never worked well in Firefox for me.

In the past few months, I've finally started working on a basic replacement NVR that works for me: https://github.com/AlbinoDrought/creamy-nvr

Like many video projects, it's a glorified ffmpeg wrapper :)

Thought it was goofy that I was still reading newsletters through my inbox. I really don't want to open my email unless I'm working. Anyways, some friends and I made Scrollz to fix that and also add some cool features to the newsletter reading experience. AI summaries, newsletter discovery, audio narrations, etc.

https://www.scrollz.co/

AI agents and testing “vibe coding”

It doesn’t feel there yet, but starting to seem some workflows could be close. And non-technical folks at business are starting to pay attention and want projects moving in those areas.

A reactive notebook with algebraic effects for building backend/AI-engineering pipelines.

Reactivity can update the state of the notebook automatically, so you don't have to keep track of which cells to execute again. Side effects are managed to make it easier to reason about while maintaining reactivity and ability to interact with the outside world.

You are a PM and Engineer - and so is the AI. You both write tickets and you both complete them to iterate on your code.

You can build webapps very quickly, especially AI-enabled ones, and deploy them on a subdomain. Other users can sign up and use your webapp, and any tokens they use will be billed to them and you will get a large cut (80%) of the margin earned on the tokens billed - as I bill 2x OpenAI API token costs to create this margin.

https://codeplusequalsai.com

So ideally you can validate your idea by rapidly building a prototype and evening earning revenue to boot.

Since I had so much trouble managing my entire digital information universe [1], I decided to scratch my itch and solve it for myself and maybe others as well. Here are my ideas about my product:

- Manages the entire range of personal (and maybe business) information/content: Documents, Media, Messages (email, instant, etc.), Contacts, Bookmarks, Calendar, etc.

- Tag based, so that where to put and find content is easy to answer. Think of a set of flat folders, on one or more devices, within which the files are stored with tags attached. Since people often find navigating/browsing files more natural than searching, virtual folders will be dynamically generated to guide navigation. Also, entire folders can be treated as atomic, and tagged/managed as one object (useful for repositories & projects). And, heuristics (and maybe AI) will be used to automatically tag files when they are imported into the tool, greatly reducing the tedium of adding tags.

- Is file based, so that all information is physically stored as individual files. This allows information to be more easily managed on a physical level: moved around, backed up, exported/imported, searched, navigated, etc. So in addition to docs, each email/instant message, contact, scheduled task/event, bookmark, etc. would ultimately be stored as a file, unlocking all the things you can do with files.

- Has a local web-based UI launched from a local agent, so actual file content does not usually need to move across the network and stays local, and the tool is also easily multi-platform, with consistent UI irrespective of platform.

- Provides a cloud web UI as well, that communicates with content devices through the local agent, so that content stored across multiple devices can be managed in one central location, even without direct access to those devices, team/org features can be provided. However, file content still stays local, except when shared.

- Provides tools for exporting data as file from the data islands of various apps and service, and backing up as files to cloud storage services.

My vision is a situation where I am in charge of my own data irrespective of whatever device, app, or service I use, can ensure that it is always available and will not be lost, and that I can easily navigate and search through it all to find whatever I want, no matter how scattered and massive it is.

[1] Here are some of my issues with personal information management affordances of current tech, which is driving me to work on a solution:

- Our data is too bound to device and vendor islands. Can't easily move my information across Apple/Google/Whatsapp, etc accounts. Can't easily merge and de-duplicate either. I almost always somehow lose data whenever I have to move to a new phone, etc.

- Hard to own your data on many services: Discord, Slack, etc. Can't easily export, search.

- Hard to have a 360 overview and handle on all your data assets and query them in a consistent manner.

- Files as a unit of information storage and management is very ergonomic; we shouldn't allow that concept to be buried by vendors for their own gain.

I am restarting my free and open source SRS kanji learning app https://shodoku.app which is based on free and open source dictionary data and Anki’s FSRS algorithm.

What I have is a basic flash card app with double sided cards (for writing (i.e. drawing) the kanji, and reading). What sets it apart is that each card contains all the relevant dictionary data, and users are encourage to bookmark a couple of words to help them remember the writing or the reading of the kanji.

What I am working on now is the database backup/sync system. I store all the user’s progress in indexeddb store in their local browser. To sync I am writing a simple patch system, so they can pick a remote somewhere (e.g. a gist on github) and push their latest patches, when syncing progress I would check the hash of the patch and apply the relevant patches.

After that I am planning on turning it into a progressive web app so users can download the app onto their devises.

https://shodoku.app/

https://github.com/runarberg/shodoku

I must confess that I have been running an experiment on HN since Q3 last year using agent bots. There have been a dozen in the wild that appears to have not been detected. Many of you have upvoted the AI generated comments. Some of you have engaged extensively with it.

It is impossible to detect them is my conclusion.

edit: I have no intention of stopping them and will continue my experiments. I do not see any ethnical issues with it as I do not seek to profiteer (although I could easily do it) and want to know how long I can keep the HN bots going.