alt Hacker NewsAsk HN: What are you working on? (November 2026)
What are you working on? What have you been curious about lately?
Comments
Hmm. A lot of smaller projects that seem to be slowly integrated into agent skills. I am admittedly moving into agent world very slowly.
Though now, with the specter of raised prices on the horizon, the local only portion of my setup gained some additional focus.
Still, from the fun mini-projects:
- attempting to replicate testing whether agents elevate inferred persona, memory, or background context above explicit user requests ( https://bloomberg.com/news/newsletters/2026-10-07/study-clau... ) - a few hackaton entries - my llm rpg got some few additions ( still nothing I am willing to publish ) - first working and useful unraid app;p
An app-only lunch bowl chain that can execute very complex customisation without breaking. We've built "cook augmentation" that takes over decision making and QA, added an app-only ordering (accessible to the agents) and ended up with a frictionless experience where you just tell your agent "I'm hungry" and we cook a fresh, hot, delicious bowl that's good for you.
Currently at prototype stage, raising angel round in order to do a live pop-up.
I've been working on a pi extension for running agent swarms. Paired with deepseek, this has been performing really well for various research tasks.
https://pi.dev/packages/agent-horde
After every run, it generates a postmortem which I then use to iterate on the protocol. I call it poor man's RSI.
I am rewriting LaTeX from scratch in zig. Byte compatible with pdfTeX XeteX and luaTeX.
It’s faster at compiling, better errors with modern DX.
I continue to work on Akariq my travel eSIM business [0]. No app or account required. I have expanded my offering to 200+ countries. In the next 1 week to so prices will come down further in Europe and Caribbean regions.
A lot of people from HN have bought travel eSIMs. I’ve received messages of support plus most people have switched over from Airalo which is expensive and generally doesn’t connect to the best network in the country.
I am building https://withthislink.com
It's a site for couples to create a wedding invitation website where guests can RSVP right on the invitation. You get a personalized link for an invitation with different domain options.
I think it looks neat and turned out rather nice.
I am trying to build a better place for people and companies in tech to find jobs and talent. Recently I redesigned the whole platform but everything is pretty much work in progress.
A site blocker extension for Chrome and Firefox. Fast and not recording any data. I was sad that I found an easy to use one or one that has no recorded my usage: https://github.com/SiGe11/Lightweight-Blocker-Extension
This one I (mostly) did without LLMs, but the other project is vibe coded, but I'm enjoying it. An RTS made by Claude in Unity: https://github.com/SiGe11/StarForge2
I am building a work/coding agent with my consulting client(s) in mind. They are a non profit and have a software (that I build) to manage property tax exemption compliance.
They have 70+ property managers as clients in the states of NC/SC. A few of the property managers are quite large, managing 10000+ units. The property managers use a mix of Entrata, Yardi, custom Excel to manage the source data. Only a part of that data is needed for tax compliance (which is what I build).
I provide help onboard the property managers. I do all my work remotely, living in India. Lately, my client (the non profit) has been creating mocks of new feature ideas and sending them to me. They use Claude desktop I think. They also maintain a Google Sheet with bugs or new feature requests. I manage the entire development and support via Claude Code.
My idea is to give my client a desktop app to directly manage certain aspects of the software. For example, we support many different Excel/CSV formats for the incoming data. These are exports from Entrata, Yardi or manually managed Excel. Our expected format is fixed but on many occasions there is data missing or poorly managed upstream. My client can easily figure out what is wrong if they get access to the files and generate tasks for coding agent which I can review. Or, if there is data missing, they can also send emails to the property managers. All property managers (my client's clients) could also be given a desktop/web app to help managing the data at source.
I would host a backend for the work/coding agent on a server. The desktop/web app would manage users, permissions and tasks between users and agents. Hopefully, this would grow into a much larger AI enabled product for property managers in a couple years.
final polishing of FinditAgent, a Windows tool which allows AI agents to research large collections of local files without building a search index or vector database - and without giving the AI any access to local file systems.
The user may ask any question, the AI will create the search strategy and queries, FinditAgent will act as a kind of Archivar, doing the searches and sending only results or utf8-excerpts to the AI. The tool offers predefined connections to the api of most cloud models - or to local running AI via ollama.
I have developed the underlying search engine for many years, now added the agent integration. It's now available on GitHub, free for private and experimental use. At this point I'm mainly polishing the documentation and testing how reliably it works on archives other than my own.
https://github.com/MHoub/Findit-Agent-Frontend
Its a one click windows install and I am very interested in hearing from people who work with large, messy document collections.
A 3D pixel-art basketball game built with JavaScript
Play: https://vladsez.github.io/iso-hoops/
Source: https://github.com/VladSez/iso-hoops
Features:
- 5 game modes (free shooting, 10 free throws, HORSE, beat the clock, 3pt contest)
- 5 courts with dynamic day/night cycles
- offline
- no backend
A friend needed to decide which cybersecurity conferences are worth going to, so I started building Find a Conference (https://app.findaconference.net).
I'm more of a UX designer / frontend dev, and I've had to learn a lot about things like postgres queues, auth and CI/CD in the process of making it. It's been a challenge trying to develop the features, while also maintaining the conference list and verifying all the info (I use some AI in the data sourcing, but want to keep it all human validated). I'd say it's just reached MVP, and it's definitely still a bit rough around the edges, but I'm already seeing some organic traffic. I'm also learning some conferences are very expensive...
Open to any feedback, especially from folks in the infosec industry!
Independent Vet - a website and app that helps pet parents find locally owned (and corporate/PE owned) veterinary practices nearby. It came out of a bad experience we had with our dog 3 years ago at a practice that had been recently acquired by private equity. The sites now get about 30k visitors a day and continuing to build in the evenings and weekends with Claude/Codex has been a lot of fun.
Feedback on your live site, typed or spoken. Each point lands on the element it's about, ready for your team or your coding agent.
I’m refactoring https://rcarmo.github.io/projects/piclaw around pi-durable (which is essentially removing overhead and duplicate code at this point, since it was already conceptually close) and wrapping Obsidian into a server-side app with a pi-durable agent to help tag and organize notes. I also have a few Go ports of things like pi-ai and core SIMD inference stuff going, slowly, since it’s too easy to get sidetracked.
As many of you, I have a few subscriptions to Claude and Codex (just me??). It's a daily occurrence to use 2-3 accounts for the same conversation (whenever the limit stops me).
What I need is something to switch the account where the usage is taken from, not the configuration.
I made something simple & transparent named Switcheroo: https://github.com/FTCHD/switcheroo
Works with other AI CLIs like Gemini, and other various CLIs like Vercel (if you're managing multiple accounts like me), Netlify, Github etc
Camera Nocta. I started making an Android app to take better black and white photos of my wife, and it snowballed into a full fledged film-styles app for both Android and iPhone (if I ever get app store approval). It's best for portraits (tuned for a lot of skin) and fashion photography.
The RAW pipeline is my own, with my own denoise, because I hate the "skin waxing" that bad noise reduction and AI seem to do. So there's lots of natural grain. The UI is somewhat opinionated because geared towards more shooting and less editing. The editing tools are my own ideas regarding how to push color (more like dye and pigment than like saturation tools).
It's out on Android, but the Test Flight has more users (1,478) than the released Android app I think. Every single artist that showed interest online used an iPhone, so I guess I should have expected that.
https://play.google.com/store/apps/details?id=com.cafesarris...
Lipik, a free open-source dictation app for Mac (https://github.com/Gourav2609/lipik). Hold fn, talk, let go, and clean text lands in any app: ums and self-corrections removed, spoken punctuation turned into real punctuation. Speech-to-text and clean-up both run on the Mac. On my 106 English test clips, lines that came out fully right went from 28% to 76% with clean-up on. Right now I'm working on Hinglish (Hindi + English in Roman letters). It gets 7 of 20 test lines fully right, so there's a long way to go.
It's a product design platform where humans and AI agents can work on the same project.
The main idea is to have a platform-independent system of record for your product that agents can interact with and use to generate code for any platform.
With agents generating more and more code, I think it's important to have a place where the actual product decisions live. Not just the design, but also accessibility, translations, behavior and so on.
I’m seeing if an agent can run a full e-commerce website. Nothing fancy, just a storefront that connects to a print on demand supplier. The agent releases new products and designs and makes sure there’s no hiccups in the orders.
Personal project - long past due to set up a home lab so I'm piecing together a SAN and some compute from old parts i have laying around and some second hand servers in a new rack I put together, along with setting up a bespoke home network instead of just using ISP stuff.
I have long been averse to configuration and customization but my needs and preferences have outgrown commodity systems and default setups, and AI is making it much easier to customize systems, so I'm embarking on this journey after over 35 years of using computers.
I am working on a casual game / experience about Indian festivals as a hobby project. This started as a Diwali night. You can walk through an old neighbourhood down to a river, light 108 oil lamps, set of crackers and send lanterns to a sky that is shared by everyone. It runs in browser without login or any ads.
I have recently added more features like a lamp you keep burning (as per Indian tradition) & a rooftop garden that keeps growoing. I have ideas to add more missions - planning to continue building on this over the year.
Some stats: Around 1,100 players have lit all 108 oil lamps (diyas) & 1600 lanterns released into the sky. Yesterday was the busiest day (because it was Diwali festival)
I'm not a code guy. I work in medicine. I've been working on a tool that can automate part of my work. If it works well it can free doctor working time and save lives too. It's related to ECG interpretation.
I'm really pleased with how my equivalent to Smart Objects turned out in Mojave Studio, my Mac image editor. In Photoshop you can transform an opaque sub-document with an affine and/or homographic transform, while in Mojave Studio you get what I call a "Smart Folder" shown inline, as if you're examining shadow DOM in web inspector. The thing you're viewing has its identity transform and the rest of the document gets the inverse transform applied, to see how it sits in context. https://mojavepaint.app/blog/smart-folders
I've always been interested in how "AI Assistants" work. There's now a ton of them out there. But you can run mine locally, and it has a couple of easter eggs ;-) as well as voice notes through telegram, connect MCP servers, and runs in a single Docker container needing less than 100MB of memory. Any issues/feature requests would be welcome:
https://github.com/mutablelogic/go-llm
No access to email/calendar directly, so that's the next thing.
I've been laid off in December, and this year I've been trying to get developers off Gitkraken and Termius subscriptions by building 2 Rust native interconnected apps.
You can check them out at: https://branchbay.dev
https://killswitch-lang.org - a programming language designed to be impossible for LLMs to read/edit/write
I’m building a personal apps platform! I still am hoping for a future where every interface doesn’t start with a chatbot.
https://github.com/momja/exhibit
You can use an agent to build small apps for you, your friends, family, etc. and use them on the platform. You can also download them and take it with you so there’s no lock-in or special sauce. Its deny-by-default security policy helps you wrangle untrusted AI code and I’m iterating on API connectors to bring your data in. This week specifically, I’ve been working on adding some UX polish to it, figuring out how I can run a reliable hosted version, but you can self host it today. I’m not a designer, so trying to learn a lot and leverage any tools I can!
Late the bandwagon but have been experimenting with agent assisted lean/formalization for the first time; it's been surprisingly high ROI vs previous fuzz coverage and has forced me to think through some interactions of cases that I hadn't yet encountered in the wild. Very useful for flushing out null behavior in my typed SQL engine [1].
On another project to inventory street trees - making progress on training a vision model for aerial tree photo detection in urban environments (original goal was to fill in central park in NYC, which is missing from their public inventory). Haven't gotten NYC to the level of precision I want, but SF has good initial results! [2]
I'm working on https://hacking.cafe a Unix/Linux environment in the browser.
I was thinking about making Bash an interpreter a library, but I just found that Vercel published Just Bash, which has the same functionality.
I'm working on Fraktur, a DJ mixing application that (also) works in the browser: https://greenvitriol.com/posts/fraktur
Working on Aqueous [1], a next-generation compressed archive format. I've just finished a long refactoring session that enables streamed processing instead of keeping intermediate results into the memory or spill file, and am now working on formats that were big enough that required this support (e.g. disk images). Meanwhile sparse file handling and in-place move and delete operation of archive members have been implemented.
I just launched a no-subscription fitness tracking app (macros, biometrics, workouts, supplements, etc.), aimed at bodybuilders. https://apps.apple.com/us/app/cheiron/id6825576174
I'm working on the interview portion of my federal income tax formalization project. I've hit my coverage target (80% of filers), so I'm doing some refactoring to optimize future development. https://github.com/emmilco/us-tax-model
Privately I've been running a "game" for the past couple months with claude where it asks me questions about movies I've seen, tries to predict my answers, I grade the answers, and then it builds a theory of my movie taste. It's gotten really good and has reached the point where it can reliably recommend movies and shows I actually enjoy.
I've been using the recent MIT weight-space programming techniques to work with neural quantum states, which are neural networks that represent a quantum wavefunction. I was able to produce a bunch of novel networks for a few hundred dollars of cloud GPU, which would have probably cost tens of thousands to train in the traditional manner.
I've been working on trying to be more helpful around the house, and trying to spend more time with my kids. I'm trying to give my partner a little bit of relief so that they can pursue their passions as well.
https://uruky.com is a private, paid, ad-free and personally customizable search engine. It was launched by my wife and I earlier this year and reached 400 MAUs and €2000 MRR last month.
https://podby.com is a simple EU static and app hosting service. It was launched last week, and is at €20 MRR.
https://fairats.com is a simple and fair Applicant Tracking System for remote-first teams, launched a couple of months ago, still at €0 MRR.
Happy to answer any questions and consider any suggestions!
My main side project at the moment is an AI workspace that's fully customizable.
I've done 80-90% of my work from inside it the past two weeks and made it open source, but haven't made much noise about it yet.
Just like the last time I posted in this thread, I’m working on The Marvelous Brass Chessplaying Automaton, a chess database/playing/opening training app for Apple platforms:
https://testflight.apple.com/join/mqfhpmHY
The original goal was an alternative to ChessBase, but with an Apple-native GUI and iCloud sync between devices, and as an improved alternative to tools like chessbook.com for operating repertoire management and spaced repetition opening memorisation. In recent months I have focused more on the chess-playing parts of the app, trying to implement human-like computer opponents at all levels, using a transformer-based model (similar to the one described in the Maia 3 paper) combined with a very small search.
Recently I’ve also been working on chess variants. Since a while, the companion Apple Watch app plays 5x5 chess. More recently, I added a 10x10 chess as an Easter egg game inside the app, and I’m now working on hexagonal chess. With the current state of the art in computer chess, it’s almost trivial to implement vastly superhuman strength in all these variants, but implementing weaker levels that are enjoyable opponents is interesting and challenging. Unlike in normal chess, there are no big databases of human games to train on.
RainbowSay https://rainbowsay.ai/
I used to watch K-drama on Netflix and pause it to copy paste subtitles in AI chat to do vocabulary breakdown, translation and grammar. So I automated it in an app which auto pauses video when it detects a sentence and gives translation, vocabulary (for every word) and grammar.
I also added same to voice from microphone, for deliberate practice speaking what was learned. To help with spaced repetition.
For about 10 months or so I've been working on this FOSS hobby project to get more practice at Linux packaging, cargo utilities, and migrating to Nix and NixOS (as well as solve some personal pain points in using a HDMI TVs and projectors with PCs).
https://tisvlay.readthedocs.io/
> TisVlay brings TV Display control for Linux, including: power, volume, backlight, AV mute, input switching plus more, with per-device profile automation via desktop session integration (idle/lock, boot/suspend, remote login/streaming, etc), and in reverse, letting the TV lock, sleep or wake the PC and its remote drive the desktop, while supporting various vendor interface protocols for multi-display setups, mitigating energy usage and OLED panel burn-in or projector bulb burn-out.
Disclosure: this project includes AI assisted coding, documentation and research, and has been a great help given physical disabilities in typing. After programming by hand well over a decade, using voice dictation scene the '90s (IBM ViaVoice or Dragon NaturallySpeaking anyone?), and worsening hands over after my CSE doctorate, this was also a bit of an individual research experiment to explore how effective LLMs could augment my personal approach to spec & test driven development.
While still beta (pre v1.0), its already daily drivable and fully documented. Being largely feature-complete, my current focus has shifted towards distro packaging, and custom ATmega firmware in Rust for HDMI CEC USB adapters to solve a few TV firmware bugs. If you'd like to give it a try, the installation guide and docs are linked below. Or if not, and you already use something else, I'd like to give a shout-out to netflector, especially for privacy-conscious home-automation users, given how common spyware is in smart TVs and projectors these days. The docs cover using it to keep displays on an isolated network.
https://github.com/netflector/netflector
https://tisvlay.readthedocs.io/en/latest/operations/isolated...
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)
working on AudioWeave, an ad free, offline first Android podcast player.
switching from iOS, I wasn't happy with the existing podcast apps, so I built this mostly for myself. I've been using it for the last six months and finally pushed a public beta.
My day to day is BE so I used the project to learn Kotlin and Jetpack Compose. Everything is strictly native using Room and Media3/ExoPlayer. I wanted a nice looking (subjective), privacy-focused app without any tracking or ads.
https://play.google.com/store/apps/details?id=com.enyodigita...
I'm building Starkeep, an open-source serverless self-hosting system that lets you keep your own data and build your own apps that run on your devices and your own AWS account.
https://github.com/starkeep-org/starkeep-core
Starkeep's common core provides file indexing, syncing, access control and more, making apps both easy to create and interoperable. Starkeep is easy to deploy, and its fully serverless architecture minimizes running costs. I'm working on an official Starkeep Photos app as well (https://github.com/starkeep-org/starkeep-apps), built for web and Android so far. Everything is MIT licensed.
Optimizing token usage for my AI coding agent: https://github.com/fluxmoot/fluximp
Its slow going. When I started this near the beginning of the year the ratio between input and cached tokens costs for the interesting open weight models was less than 10 and provider side cache hit rates weren't very good, so it wasn't expensive to break the prefix to remove old things from the context and do rolling summarization to keep the prompt size down.
Now some DS v4.1 Flash providers have a ratio of 33 to 50 and >99% cache hit rates. That makes the cost of prefix cache breaks a lot higher and its definitely harder to get ahead on total cost* of the usual approaches to context management for DS v4.1 Flash specifically.
* - the number of requests is reduced by a factor of ~3 relative to oh-my-pi on the same benchmark; the task is long enough to require compaction once when using oh-my-pi
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.
I'm bootstrapping a Dyson sphere https://muni.works/research/vision
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
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.
I got tired of always having to update my wife, and my family members résumés. So I created https://theresumechef.com It's private,free to use, no pay walls or monthly subscriptions.It has10 resume templates to choose from to get you started.The final resume can be exported as a PDF, or stored on Google Drive or Dropbox in JSON format. There are still a few issues, but I've learned a ton by just creating and launching this project. I have some exciting new product ideas, but I'm currently waiting on my Dutch KvK number to accept online payments with Stripe.