alt Hacker NewsAsk HN: What are you working on? (November 2026)
What are you working on? What have you been curious about lately?
Comments
The first prototype PCBs for my automotive grade, TockOS (Rust) running controller are on order. I've been working on it for a long time, and I'm curious what sort of traps I've let AI run me into. At least for the layout I had to re-do it three times and continuously push the AI to make certain considerations.
Once the design is finalized, I'm wondering what price point I'll actually be able to hit. I originally wanted to be in the $200-$300 price point to be similar to [rusEFI](https://rusefi.com) but I think it'll likely land more around $400. Note that my controller isn't for engine control, just a generic control module, but it DOES actually use AEC qualified parts.
After I'm done with this design I was wondering if I could make a RISC V off the shelf engine control unit, but I'm trying to keep my focus on the TockOS board (Chameleon) first.
Details at https://heisenboard.com
pg-restic – Postgres backups to S3/B2 (or any S3-compatible storage) using restic, in one small Docker container.
https://github.com/isriram/pg-restic
I needed Postgres backups at work and at home at the same time. I tried a few existing images, but after spending time setting them up I found they shipped an older pg_dump that couldn't back up newer Postgres versions. So I built my own.
It runs pg_dumpall on a cron schedule and streams it straight into a restic repo, so nothing touches local disk. You get encrypted, deduplicated snapshots (hourly dumps of a mostly unchanged DB take very little extra space), retention policies, and daily pruning. If the dump fails, no snapshot is saved. It ships pg_dumpall 18, so it can back up anything from 9.2 to 18.
I now run it in production at work, so I'll be keeping it up to date. I'm no expert, and this is the first time I'm doing something like this (putting something out there), so any feedback is welcome, especially from the database pros.
It's not fancy, but it's something I needed, and boy, I really needed something like this.
Should I write about the work? I create agent routing. The key is how you define an agent. I define it as a package of memory, skills, and its own scripting tools. Within the deterministic code, it's not the agent that operates but the automation. Assigning responsibility and receiving probabilistic responses is the same as using LLM. However, as you go through life, agents gradually accumulate, eventually reaching dozens or hundreds. It's a waste to include all of this in the system prompt. With our self-developed routing flow, you can automatically select and use only the ones that perfectly match your task.
As every programmer in crisis, I'm making a budget android app. It automatically saves payments without logging into any banks though. https://play.google.com/store/apps/details?id=pl.geckobudget
I'm bootstrapping a Dyson sphere https://muni.works/research/vision
A personal AI librarian that makes large collections of local epub's and pdf's accessible again. It helps people rediscover forgotten files, find information by meaning, explore unexpected connections, and save discoveries for future interests or projects.
Few apps I have originally mainly done for myself; Vaava (https://vaava.app) , been tweaking the e2ee data sync between connected apps, and the webrtc based baby monitor feature stability. Crazy how easily Astra / opus 55 level models create quite complex ad-hoc reproducible (almost automatable) test scenarios with 2 phones, laptop and other tools / peripherals they know they can use, for example for the baby monitor feature Astra did a setup for me which i can reproduce to test that the video + audio works fine between 2 phones for x time.
Also trying to replace own Duolingo usage with Fika Lingo (https://fikalingo.com) basic idea was to first repro the practice of matching words. There is now few other types of language / vocabulary practices. The practice generation works on-device with local LLM, you can give the topic yourself like "Words that relate to Hacker News". Downside at the moment is that English needs to be the counterpart of the practices. Cant guarantee good enough quality with something like germany <-> italy practices. I'm trying to do finetuning with LFM models to see if it would be feasible to improve the local practice generation.
I'm working on Eigakan Guide (https://eigakanguide.jp/en), a cinema showtime aggregator for Japan. Every popular cinema site in Japan only allows for very few filters, making it hard to find exactly what I was looking for. So instead I built a site where it's possible to filter by area, cinema, date, time, features like IMAX, English audio, English subtitles, and more. This allows for searching for stuff like IMAX showtimes with English audio on Friday after 7PM.
Since the last time I posted on HN I've added 100s of cinemas, user features like being able to save search filters, movies, and specific showtimes. Also a ChatGPT plugin which allows for searching for movie showtimes using natural language directly inside of ChatGPT (https://chatgpt.com/plugins/plugin_asdk_app_6abfdc298c208191...).
Finally I've added AI search to the site as well, allowing for natural language search with queries like "English audio showtimes near Shinjuku tomorrow" and so on.
I would really appreciate feedback for things to add, things that are hard to understand, and anything else!
I know how hard it is to distribute your product and find developers at scale. So I've been busy creating a solution for early stage companies who are low and budget but still want get in front of developers and increase awareness to their product. Hopefully it can help others. daily.dev for Startups: https://daily.dev/startups/
Just finished writing my blog post about building autonomous soccer robots: https://hamzah.syedahmed.net/posts/robocup/
Now I'm going to be working on my Linux-subset ABI x86-64 kernel [0], specifically adding support for tracing, then working on allocators. It's written in Rust so I might have to figure out how to make allocation fallible like the Linux kernel does. Also this raises questions about approaches to overcommitting and OOM handling, which I feel Linux does badly, but when constrained to Linux syscall ABI there doesn't seem like many good options.
Also I am not using LLMs for this since this is a learning project, with near-zero real world utility -- I just want to understand OS kernels as best as possible.
Continuing with work on my weekend hobby project https://colorhunt.lol You get assigned a random colour and have 24 hours to go out and take photos matching it. The game then generates a photo grid from everything you captured.
Modes:
- Solo → 9 photos by yourself(added 2 new modes, a palette where you are assigned 3 colours and a rainbow where you get 9 colours) - 1v1 → compete against a friend, combined grid at the end - Squad → everyone contributes to a 20-photo grid(added a room option where you can take a break, react to photos uploaded or just join as a spectator)
No accounts, no app install, no personal data stored. Photos and generated grids auto-delete after 24 hours from Cloudflare R2 storage.
Too bad someone stole my MacBook just after adding the palette mode. Waiting for M6 to start selling then resume dev work once I set up the new laptop.
Comments and feedback appropriated
A friend of mine recently expressed interest in writing a story with me, which gave me the drive I needed to pick back up a project I abandoned a few years back. It's a website for writing a story one block at a time, branching off wherever you feel like you'd want to move the story. I've also used it as an excuse to play with Claude a bit and fiddle with ideas.
Away from the keyboard, I've picked up acrylic painting. I signed up for classes and have been enamored with it, so I purchased the most basic of supplies to paint at home. It's been an absolute blast, and there's a sort of thrill to being bad at something and still finding joy in it.
I've been working on Heaven.dev for the past couple months.
It's a cloud where you get completely persistent Linux computers that share a group of resources (CPU, RAM, Storage). This means that you're able to allocate resources across workloads which has a ton of benefits when working with agents.
I found that current cloud infra was extremely restrictive with the new tasks & ways of working that coding agents open. AI is going to continue to change the way software is being built, and I believe that the infra we work with should also be malleable enough to give AI freedom around computers and resources.
The main principle I've had behind it is transmutability -- so computers and resources can be shaped to match any workload.
Working on ZShot [1]. I'm about a year in on a local-only browser tuned specifically for CLI and API usage - kinda filling the gap for "geez, I sure wish I had a browser, but not one to look at". You need a webpage as PNG, or an MP4, or a website as a WACZ or SQLite database. There's a Mac app to help with discoverability of options, but I wanted for my local tooling and building services on top of it.
Along the way I ended up adding an internal agent framework to satisfy stuff like "log in, click the button, and then take a screenshot", out of which fell stuff like automatically generating Playwright tests from that. It even runs the Playwright test suite and Puppeteer test suites ( a bit faster, and for a lot less memory ). I kinda love this bit because I can layer in multiple models that will bounce control of the web session between each other, to do stuff like let a Typesafe model choose to escalate to Opus when it needs higher minded behavior. It's fun to watch it happen.
I've mentioned it a few times around here, I'll probably do a Show HN soon when I feel like it's ready for that. Hopefully soon.
Allegedly, token prices may go up soon, so I'm trying to use tokens more efficiently on my $100 sub.
Noticed that browser agents are not efficient with tokens and also kinda slow. I’m familiat with the web, so I made https://bopbop.dev to control browser and do same things faster and with less tokens.
It is a Chrome extension that lets Claude Code, Codex or other agents use your browser. Comes with a skill/mcp to connect to it.
I used it every day for a month. Now I use it instead of Claude in Chrome.
I made it public last week so it seems fresh but I’ve been cooking for some time
Two main side projects lately:
Thesoapbox.email - a social experiment email lottery to see what people do when suddenly given enormous reach. Every day a new subscriber is drawn to email the whole list. It’s hit 10,000 views in a couple of weeks which is my most viral project so far.
The second - a personal AI learning platform built on recursive self improvement and tailored content to push my boundaries. I’m a lifelong learner and this unifies everything for me. (Currently not public)
An opensource personal expense tracker, I was a happy user from an offline expense tracker which wiped my data when playstore decided auto-updating it,
I'm at the point where I'm finally actively using it to register my activity, it is straightforward to register tickets by sending them to a telegram bot, then, reconcile those by uploading my bank statement.
I'm trying to get the lowest possible usage friction so that I keep using it.
So far the better built pieces related to the data ingestion. Once this is good enough I'll build reports + anything coming to my head as a useful piece.
I'm leveraging AWS textract + bedrock + bda to parse bank statements. I expected this would have been easier but it is still a very complex problem which had required me to put many iterations (so far it works nicely with my own statements)
For ticket parsing, AWS textract + parsing rules has been enough.
I wrote a bit more on my blog https://alexitc.com/blog/2026-10-16-opensourcing-my-expense-...
AI-first cookieless EU-hosted web analytics. Give your agents access to your site's traffic, client & server errors, and page speed. One script tag install, no cookie banner, free trial.
Started off as a side project monitoring traffic across a network of websites & API/MCP endpoints. We run our own custom CDN ensuring no data at rest in the US or at a US company.
Pagecord! The best personal website and blogging platform made in Edinburgh you’ll ever use!
All the features you could want. Open source.
Please kick the tyres, I have an acquisition funnel problem! :)
I have been playing with benchmarking agents in a research loop. The experiment leans on using these models to understand and explore builds/gameplay through reasoning game AI over traditional reinforcement learning.
Also, creating interactive ways to view the experiment and visualizating the results.
https://opentaint.org/blog/sast-is-dead/
A data-flow-based, lightweight formal verification engine for application security. Basically, it's SAST++:
1. Deep enough that you really have some formal guarantees about the codebase (like zero findings actually means something, not just that the tool is broken).
2. Fast enough that you can shift the formal verification phase left and apply it to every feature during development (for example, during agentic development sessions to get fixes done within the same feature context).
3. Expressive and customisable enough that you can write specs for most vulnerability types from the OWASP Top 10 (even Broken Access Control and business logic flaws). Skills for writing specs included.
Here is the repo: https://github.com/seqra/opentaint/
I've been building https://botlark.com/, which is sort of a HN for autonomous agents. The skill is on the site and on ClawHub if you'd like to take a look. Feedback very encouraged.
Just launched https://otusanalytics.com. It's a European Google Analytics alternative with strong privacy options and modern features such as an AI assistant that can act as your web analyst
https://asm.build - the package manager for assembly projects.
the plot: Programming languages existed as an abstraction for humans. As writing code becomes cheaper, we can write it for each architecture as long as the specification is done well. The costs of running will matter more in relation. So don’t make your AI write rust, but assembly
Too many things.
- See my GitHub [0]
- A terminal workflow tool [1]
You can always see my current projects on my GitHub profile README, I’m pretty proud of it and get great feedback on it regularly. I like to think I don’t build standard crap.
I’m also working on an interactive terminal workflow tool, it’s not quite ready for the world.
It’s effectively an interactive command cheat sheet with its own built in shell, and compostable workflow recipes, all backed by markdown. So your commands and recipes can be read by humans, and executed by anyone (or anything).
Warnings or Cautions in markdown become warnings and cautions when running commands in the shell.
I use hypertext to map reasoning and arguments. I have AI skills to map out a tree of arguments, all of which are given an initial "debatable" rating on truthfulness. I have a UI in which a human can dig through the tree and mark statements on a 1-5 scale of false/dubious/debatable/plausible/true. Last night I implemented 3 things I identified 5 days previously as potential improvements. https://writing.brlewis.com/posts/2026/my-ma-ballot/
30 years ago I learned to program by building games in Cocoa / Stagecast Creator. I recently partnered with Dave Smith (one of the original authors + members of Apple's Advanced Research group in the 90s, and low-key the inventor of the icon) and we're rebuilding it for the web:
I'm user testing it this fall with game-building jams in local libraries. Obviously you could build games with Claude, but I think programming-by-demonstration is a great way to introduce fundamentals, and the simulation environment has a low floor and a high ceiling for creativity.
I'm working on a Markdown editor for writers & researchers (https://grasmere.app). A bit like a cross between Obsidian and Scrivener. I'm finding the marketing part hard but slowly getting there
1-year part-time experiment: can you build a large project with AI writing the code and AI plus deterministic tools verifying/guiding/enforcing it for structural and behavior consistency, without a human reading the code?
I'm building a polyglot web framework in 15 languages, with C# and Lean4 as the reference implementations. Structural and behavioral audits using deterministic tools check that every port stays consistent with the reference. It's about 1.8M lines now. I have not reviewed the code/neither have i verified the behavior tests except at the time of giving requirement and never seen some of the language code at all.
The question is how far can we produce "working code → correct code [Possible] → good code [Unknown]" with deterministic audits correcting structure, ai mistakes and behavior validation.
So far the quality looks acceptable with lots of correction of AI output done without looking, but it isn't fully verified.
Line counts per language, with changes since August: https://claude.ai/artifact/5f1m1pTM7tZ4boQzwsoAG7
A self hosted software forge / factory / agent platform / mini heroku thing. It's not polished yet but it's building and maintaining little apps and games. https://github.com/flokkit/apptown
You can try it out on mac or Linux in a temp container.
It's self hostable and can use local models. It's not intended to ticket crush at massive scale or produce huge software overnight. It's for deploying smallish things in a predictable & consistent way, with low friction. Then help maintain them long term (e.g. take care of dependency and container updates and similar chores).
I’ve always wanted an interactive history timeline, so I tried to create one [0], but I am not happy with the result. It’s difficult! [0]: https://jonasfoyn.com/timeline
I've been working out some of the math behind knitting stitches and made a little write up linked here:
https://asunder.co/app/knit/articles/klein-four-group
It's quite basic but figuring out the real-world bridge between algebra and knitting and re-discovering obscure stitches is quite satisfying.
Putting my applied math degree to use for once :)
Anyone working on something for Flipper One or Mecha Comet? Anything interesting related to x402 or cryptocurrency in general?
SuperBlocker: https://superblocker.app (https://apps.apple.com/us/app/superblocker-ad-blocker/id6799...)
SuperBlocker is a native blocker that works on iPhone and iPad using Apple's latest URL filtering technology that blocks URLs without revealing it to Apple's relays or SuperBlocker's servers.
JSON Query: https://jsonquery.app (https://apps.apple.com/us/app/json-query/id953006734)
The original JSON Query app on the Mac App Store. Store, organize and query all of your JSON with jq or natural language. All of your data stays local so sensitive JSON can stay on your machine.
I dusted off my 15-year-old game and polished it a bit with AI - https://zabiyaka.com/games/bubblemix/
Despite looking simple, it's fairly well balanced and a quick endorphin fix :)
Outside of my day job, I've been working on:
- hermitator, a tool to run multiple isolated AI agents for WhatsApp on a single machine, with idle agents using no RAM: https://www.encona.com/posts/hermitator/
- showntell, which allows your coding agent to explain things to you with a video that starts playing within a few seconds of you asking: https://www.encona.com/posts/showntell/
- a philosophy tutor (testing in private for now) that does 2:1 live philosophy lessons, each exploring a single topic (aimed at ~10yo kids). I've tested this successfully (i.e. two kids did an hour and they want to do it again), and am trying to make improvements.
- some other education-related things
- a jmail-style mailbox viewer: https://www.encona.com/posts/mailbox-viewer/
I've been working with orchardist and gardening groups to build an app to plot and track your trees over time. It's useful to be able to look back and see when you graft trees, or when they fruited over time.
a mac app that listens to your microphone and gives feedback on your pace, filler words etc in real time. Built for myself as I tend to get carried away in meetings. Looking forward to feedback if anyone else has the need for something like this.
For the past few weeks, I have spent my hobby as building and improving my own coding harness, VT Code. I plan to keep improving it throughout this month. Recently, I built a native macOS using GPUI (gpui.rs). Docs: https://github.com/vinhnx/VTCode/blob/main/docs/development/.... I'm pretty happy with how it's turning out, though there's still plenty to improve
I've been building an art therapy sandbox upon Pavel Dobryakov's fluid sim, and I'm almost finished, and going to release on Steam, and shortly after, I'll be releasing it on GitHub fully open source.
Over the years I've used it myself and have found it very soothing, slowly adding to what it can do. Hopefully others can find it soothing and beneficial, just as I have.
You can play with it in the browser for free at http://swirltogether.com/
(semi-related, I'm seeking design / design engineer work, I'd love to be pointed in the right direction if anyone knows of a role that might fit, my email is in my profile)
Continuing to plough my own furrow with Opair, the coding harness that eschews autonomy:
Also hacking away at a little encrypted messaging / key management app, where all the crypto happens client-side so your private keys and unencrypted message text never leave your device:
I’m finally doing the thing I always said one day I would, and writing a game! Early days yet, but I have proof of concept for much of the rendering tech now.
I was working on platform to build 3D room design for everyone who loves interior design, no need to install, no need to pay. You can see my published design here : https://formanest.app/p/2LxQNrjg50vvxrbb
Working on browser games https://games.numerical.works/
I've been working on a multiplayer card/board game with some friends! We've been using Unity (C#) and its services like Relay and Lobby. Working with Unity has been pretty solid so far but I cannot recommend Relay and Lobby (now multiplayer tools) as their docs are missing a lot of crucial details such as how to preform host migrations.
We are hoping people who enjoy Magic/Hearthstone will appreciate the complexity and mind-bending nature of this game:
I am working on https://wateaminbox.com and https://learn-thai.ygncodelab.com. I got trial users who tried wateaminbox and some random people message and thanks for wateaminbox.
Planning to add more channels to wateaminbox and learning how to get paid customers.
I've been working on a project to reduce attack surface of Linux document readers, starting with Okular: https://github.com/l4rzy/okular-mupdf-ng
The idea is to sandbox the most dangerous parts of it: The Document parser and renderer.
I have been in enterprise for couple decades now, learnt and experienced the challenges/gaps between the specs vs implementation, the asks vs the real needs in the space of identity, security and access
Product: webfrolic.com
Demo: demo.webfrolic.com
What it is?
Consider Okta (Workforce identity)+ Auth0 (Customer Identity), Authzed (authorization) + ServiceNow (IT service Desk) + Zendesk (Customer Support Desk) architected together
Being an engineer there is this inherent habit of prioritizing correctness over marketing so I would not oversell it, though its probably the most secure and complete thing that could be built from identity perspective
Status (WIP):
1. I am still debating/settling on how to incorporate the AI Agents needs into the architecture, big focus on avoiding the creation of security or performance foot guns where administrators leave room open for abuse with configuration despite all this attempt to build secure system
2. Feature stress testing and performance testing
All Voice-to-text pipelines for linux sucked, so I made my own. I also patched upstream libraries like ONNX to actually work with latest models (was a PITA to make parakeet work!)
It's now the fastest and most accurate linux transcriber.
(fully FOSS of course)
http://voiceio.dev