alt Hacker NewsAsk HN: What Are You Working On? (December 2025)
What are you working on? Any new ideas that you're thinking about?
Comments
This is actually pretty much as done as it's going to be (could use some nicer UI feedback, i.e. how you actually use the app) - it is actually just a demo for an effort I undertook to mod Datastar to support nested web components. I am writing it up as we speak!
Instructions: you have to answer three questions; each one will auto-submit once your response goes over 100 characters; the answer to the third question is your "post". It's a proof of concept of a friction intervention for social media to encourage slow thinking before posting (and hopefully reframing negative experiences in the mind, it's kind of dual purpose).
SecurityBot.dev is an all-in-one uptime, performance, security, and SEO monitoring tool. I launched it a few months ago and have been iterating on it ever since. Later this week SecurityBot.dev will log its 1 millionth uptime check which is pretty cool to see.
It includes the usual uptime monitoring service that you see everywhere else, but also features such as a PageSpeed Insights monitor (https://securitybot.dev/pagespeed-insights) and a broken link checker (https://securitybot.dev/broken-link-checker). I continue adding new monitor types as I personally need them (and also based on use feedback).
I am working on two things.
The first is a customizable digital math workbook. Currently the demo covers fourth grade math. There is a practice mode where you can select the skills you to want practice. There is also a customizable dashboard where you can setup your own widgets to practice math skills in different ways. I am working on some pre-made dashboards to help users get started. The next plan is to cover fifth grade math skills. My plan is to cover first grade math up to Calculus and High School Physics. I envision it as a companion tool for Khan Academy/Math Class/Math Books. Check out the demo. No signup required. Progress is only stored locally.
The second thing I am working on is an application to practice Cangjie. It's a Chinese input method that has been around for a long time. It is based on a visual decomposition of characters. Each character is represented by one to five codes and the majority are unique. My application teaches Cangjie like keyboarding (QWERTY) is taught to young students. You learn the location of the keys, then some basic words, then start typing sentences. I also have a free demo for it as well.
https://demo.cangjieworkbook.com
Feedback on either project would be appreciated.
Wanted to save up a few tokens when passing data to LLMs and did not like anything on the market, so I made minemizer.
Minemizer is a data formatter that produces csv-like output, but supports nested and sparse data, is human readable and super simple.
It produces even less tokens than csv for flat data, due to most tokenizers better tokenizing full words that contain a space before the word, and leads to less fragmentation.
There are many cool things I discovered while running tons of testing and benchmarking, but it's getting late here.
Code, benchmarks, tokenization examples and everything else can be found in the repo, but it is still very WIP: https://github.com/ashirviskas/minemizer
Or here: https://ashirviskas.github.io
EDIT: Ignore latency timings and token counts in "LLM Accuracy Summary" in benchmarks as different size datasets were used to generate accuacy numbers while I was running tons of experiments. For accurate compression numbers see compression benchmarks results. Or each benchmark one by one.
I will eventually fix all the benchmark numbers to be representative.
Working on Bold (https://boldvideo.com)
Video intelligence platform for coaching programs and training companies. The problem: these businesses sit on 200-500+ hours of video content that becomes a "content graveyard" - students can't find what they need, coaches burn out answering the same questions, churn stays high.
We do deep transcript + metadata extraction, then layer RAG search and an AI assistant that can answer questions with timestamped citations back to the exact video moment. Think "ChatGPT for your video library" but with accurate sources instead of hallucinations. Tech: Phoenix/Elixir backend, Next.js portals, two-tier RAG architecture.
Currently serving a few coaching programs in high-touch sales mode. Would love feedback from anyone who's built RAG systems over media content - curious how others handle the signal extraction problem (transcripts are noisy, you need to identify what's actually being taught vs filler).
I'm in crunch mode doing the internationalization and localization of my spreadsheet engine. This is a rabbit hole and a nightmare in Excel, so a big opportunity for us to get this right.
This is the PR: https://github.com/ironcalc/IronCalc/pull/616
Feel free to comment and destroy it!
You can test it in: https://testing.ironcalc.com
I'm working on Tech Talks Weekly which is a free weekly email with all the recently published Software Engineering podcasts and conference talks in the past 7 days.
Every week I pull all the new talk recordings from hundreds of conferences (Devoxx, KubeCon, PyCon, QCon, LeadDev, JSNation, and many more) and even more podcasts podcasts. I feature the ones I think are must-watch with short summaries written by me, then include a list of everything else uploaded that week.
It started as a personal project to fix my own messy YT subscriptions and RSS feeds and now 7,500+ people read it.
I also publish extra editions from time to time like “The Most Watched Talks of 2024” which made it to the HN front page.
If you watch software engineering conference talks or listen to podcasts, you might find it useful.
I’d love to know what you think!
I've spent a considerable amount of my free time over the last few years working on an open-source game engine for making Zelda-like games (link in profile). It's been around for a few decades, and I played it when I was a kid - and now I'm contributing heavily to it. To give a since of scale, there's ~1000 custom fangames made in this: so pretty niche, but if your thing is Zelda it's got some real gems.
Most of my time has been spent practically rewriting the engine from just single-screen play areas (like Zelda 1) to be free-scrolling (like Zelda 3). I've also put lots of work into supporting all platforms (was just Windows; now it's also Mac/Linux/Web). And I've delved into tons of interesting programming projects while working on this: a deterministic record + replay testing system; a garbage collector for our custom scripting language; JIT compilers for x64 + WASM; a VS Code language server; the list goes on...
Anyhow, this month I'm trying to polish it up as much as I can so we can officially release the next major version.
Building Arivu: CLI/library that normalizes fetch/search across a bunch of sources (arXiv, PubMed, HN, GitHub, Reddit, YouTube transcripts, RSS, web pages…).
I use it as a context fetcher i.e grab an abstract/transcript/thread as clean text/JSON, pipe it into summaries or scripts.
Also runs as an MCP server (experimental), so tools like Claude Desktop or CLI assistants can call the connectors directly.
arivu fetch hn:38500000
arivu fetch PMID:12345678
arivu fetch https://arxiv.org/abs/2301.07041
Analyzing frontier LLM performance on my favorite daily puzzle game (https://www.nicksypteras.com/blog/cbs-benchmark.html) Next step is to assess how well the LLMs can create their own new, logically satisfiable puzzles in the same style. Then I'll have them battle it out, with one model creating a puzzle and the other attempting to solve it!
Working on https://usedigest.com — a personalized daily digest.
It aggregates data from across the web into a single feed, pulling in news, weather, newsletters, social posts, Reddit, YouTube, and more.
I also finally launched my first iOS app that goes a step further. During onboarding, you set your preferences once. From there, AI automatically prepares your daily digest for you. Each morning, you get a notification when it’s ready, with everything relevant for the day ahead: meetings, weather, health data, commute insights, and the news you actually care about.
I'm trying to build a native postman alternative using Rust + Iced. I want it to use .http files as its collections and .env files as its environments. So that data is stored in plain text and easily editable by AI and usable by other apps like VSCode rest client.
I just finished https://www.lexaloffle.com/bbs/?pid=178450#p , my Signalis themed version of the Regicide card game.
It's mainly a distraction from enterprise programming, but it does have some parts that might be interesting to Lua programmers, like automated test suits, functional programming point free style and deploying to a raspberrypi via justfile.
The git README kinda doubles as a blog post: https://gitlab.com/michaelzinn/replicide
Back in 2019 I created a Google Spreadsheet titled "family debts" that allows my family (4 siblings and my parents) to record when we owe each other money, and periodically settle up. I later learned that I recreated Splitwise, but having something like this with trusted folks has been hugely useful. We have over thousand entries, and use it constantly for splitting gifts, buying something at the store for someone, etc.
Om Friday after Thanksgiving I spent half a day building a telegram bot that accepts an address and a list of Amazon links, and in turn orders the item (at a discount since it uses my Amazon credit card), and adds it to the above "family debts" spreadsheet.
I really like the idea of programmable, trusted lending like this, and feel like it could be extended to other groups that you implicitly trust.
working on bringing some basic, banal bits of infrastructure management to real-world traffic issues.
I'm literally trying to fix broken junctions around me.
It's at the same time laughably easy, and wildly complicated.
I'm calling the alternative, correct junction a 'traffic bean':
https://josh.works/traffic-bean
It's relevant to software, sorta. I've got rather a lot of GIS/mobility-related data available here. It's just a rails app that renders a bunch of my strava activity data all at once: https://josh.works/mobility-data
The fixes are entirely accomplishable with nothing more high-tech than traffic cones. They can be upgraded to more permanent and pretty physical objects, but the key bit of the traffic bean finds traffic cones fully sufficient. No half-million USD traffic signals, no red/green/yellow light cycles. continuous flow. safety. peace.
Some stuff that's obvious in some domains, like "at high-throughput times, don't allow key bits of infrastructure be completely unusable".
Bringing this to american municipalities is like trying to speak a language with someone that doesn't speak your language, but demands that you treat them as if they do.
it's been a big, long-running project. Most tradition in the USA is really a fig leaf for supremacy, and people can smell that I'm coming for their supremacy a mile away, and they immediately begin deploying emotional defenses.
Or so it seems.
https://github.com/sangeet01/limitnumen
As aspiring young man learning ai, I have successfully solved the theoretical limitation of embedding using hashing. Now working on turning that to RAG system as have solved the retrieval and now wanna complete the system.
Cheers :)
I'm working on a hardware/software utility to play Switch/Switch 2 games remotely with my brothers. I found a way to emulate a Switch Pro Controller using a Raspberry Pi Pico based on several different sources (look in the README for more info). I used that to write a firmware for the Pico (with the help of GPT Codex 5.1).
Then I wrote a Python program that connects whatever controller my brothers want to use (as long as it's supported by SDL2.0) and forwards that data from their computer, through Parsec, through a USB-UART adapter, to the Pico, then to the Switch. I then have a low latency capture card (Magewell Pro Dual HDMI I got off of ebay for $100) forwarding the video and audio from the Switch to my PC which I share to my brothers via Parsec. The audio was a bit tricky to get right, and ended up having to use a Virtual audio cable and Voicemeeter potato (a software audio mixer) so that both myself and my brothers could hear the audio.
It works surprisingly well and the latency is pretty low. I even got rumble working! (but not motion controls. If anyone wants to attempt it, I will accept PRs). I haven't done any formal benchmarking for performance, but my brothers and I were able to play Smash Ultimate without too much bother about latency.
You could also use the accessory Python library I made to automate switch controller presses (look in the examples directory). Might be useful for TAS speedruns?
The project is here for anyone interested. It's a bit rough and needs some cleanup and maybe a video tutorial on remote setup. But here is the WIP:
*Supex* - Agentic coding for SketchUp.
Working on a house renovation project in SketchUp, I wanted the same workflow I use with Claude Code: describe what I need in natural language, let AI write and execute the code, iterate quickly.
So I built a bridge. Python MCP driver talks to a Ruby extension inside SketchUp via JSON-RPC. Claude Code can now write Ruby scripts, execute them directly in SketchUp, take screenshots to verify results, and introspect the model - all without leaving the conversation.
Still very early (macOS only, requires SketchUp 2026), but it's already useful for repetitive tasks and parametric designs. "Create a spiral staircase with 15 steps at 18cm rise" is more fun than drawing it manually.
https://github.com/darwin/supex https://github.com/darwin/supex/tree/example-simple-table
We want to speed up adoption of custom AI, but most people suck at building it (no expertise, money, time, etc.).
We thought, what if you could "Vibe ML" your way to it? Allow any AI engineer or PM to build custom AI directly from their current implementation.
So we built these agents that orchestrate the entire life-cycle of custom AI. We start by hooking into how you use AI, prepare/label your data, detect the best recipes for your task, fine-tune, and deploy it for you. Really tried to simplify the entire process.
We aren't entirely sure about the UX/UI patterns. We aren't going chat first because if most people don't know where to start with ML, how in the world are they going to prompt it!?! Instead, we auto detect the AI tasks you've built and go from there.
I’m still working on MLSync.io, an ETL platform designed to remove the technical friction of synchronizing MLS data using RESO protocol and provide simple access to the MLS data via SQL and REST API.
And as everyone now, I'm experimenting with LLMs to bring some new AI-related features to the service.
On another project, we've now beta testing (in ordination) Asus GX10 processing power running on-device LLMs for _local_ processing of patient medical data for 'differential diagnoses, implant plans and risk profiles in real time while the patient is in still in the chair'.
I've continued to work on an open-source containerized agent framework called Capsule Agents. Lots of progress this month and I have my first one live in my homelab! I'm training for a marathon and its able to send a discord message every morning with the workout on my schedule and the "why" behind it.
Here's the elevator pitch for the framework:
Its built around 3 key ideas I've dealt with inside the agent ecosystem 1. Agents become far more capable when they have access to a CLI and can create or reuse scripts, instead of relying solely on MCP.
2. Multi-agent setups are often overvalued as “expert personas” but they’re incredibly effective for managing context, A2A is the future.
3. Agents are useful for more than just writing code. They should be easy for non-engineers to create and capable of providing value in many domains beyond software development.
If that sounds interesting take a look! https://github.com/brycewcole/capsule-agents
I've built this website helper to browse these "What are you working on" comments: http://waywo.eamag.me/
Updated manually so expect some delay :)
Trivyn: Ontology-first knowledge platform. Runs on a single machine, via a single executable. I wanted a simpler alternative to the large complicated enterprise products that tend to dominate this space.
I'm really trying to get a private beta out the door by Christmas. I do plan to have a free version for academic/personal use.
Backend is written in Rust, uses oxigraph for its triple store.
I’m working on a deterministic execution layer for AI systems.
The idea is to treat LLMs as constrained components inside explicitly defined workflows: strict input/output schemas, validated DAGs, clear failure modes, and replayable execution. Most “AI unreliability” I’ve seen isn’t model-related — it comes from ambiguous structure and hidden assumptions around the model.
We’re exploring this through a project called FACET, focused on making AI behavior testable, debuggable, and reproducible in the same way we expect from other parts of a system.
Still early, but the goal is simple: less magic, more contracts.
I just built myself a DYI TRMNL (https://taoofmac.com/space/reviews/2025/12/13/2200), decided to hack together my own server for it, and (serendipitously) got asked if it could be an official contribution: https://github.com/usetrmnl/byos_fastapi
Then I decided to hack my own ZigBee power meter (to keep track of my meter’s LED pulses) and fought with CMake for eight hours straight, because embedded.
It was a nice weekend.
https://monofocus.lovable.app/
Yup, it's another task manager.
I made it for myself to help me focus on one task at a time, hence the name.
It implements my number one productivity hack of picking a task and setting a timer. Time spent on a task increments.
Data is stored locally in the browser although there is a sync option i wouldn't shake a stick at if I hadn't built this myself.
Plus it's a PWA! Those are lovely.
It's feed aggregator backed by a web crawler that tries to find interesting RSS feeds. Posts are sorted by inverse frequency with the hope that time between posts will serve as a good proxy for quality.
I've been having fun with it! The results are a little strange, sometimes, but I've found some interesting sites that I never would have found otherwise.
Making music with the machine!
Album here: https://open.spotify.com/album/3e6k9eiGUlOBcoI2yd3DrM
Written about the process here: https://songxytr.substack.com/p/on-making-music-with-the-mac...
A prompt engineering tool that takes vague prompts and transforms them into context-rich JSON/XML structured prompts. Fully customizable and tracks prompting history safely in tab session, automatically injecting context and learns style.
Makes outputs of any AI so much better due to restructuring and breaking down requests into instructions AI can easily execute upon and mitigating risk of hallucinations. Perfect for complex tasks like coding and content creation.
Exists as a free chrome extension right now. Would love if you tried it and have any feedback!
Email me at [email protected]
https://codepathfinder.dev Currently working on AI-Native Static code analysis and currently it's open-source.
I’ve been working on a TLS proxy/TLS terminator that can handle 3000 TLS handshakes per second (basically an stunnel replacement, but stunnel crashes at under 100 handshakes per second) as a pet project, but I’ve realized that with some polishing this can be really useful.
https://github.com/novotimo/tlsproxy
This is still in development (todo are privilege dropping, in place config reloads, log burst suppression, multiple listen sockets (which paired with the Linux kernel gives free load balancing capabilities), and detailed TLS configurability), but it already matches both nginx and HAProxy’s speed (entirely bottlenecked by OpenSSL crypto by this point) at a tiny fraction of the attack surface and memory footprint (10-15kb per worker process last time I checked).
If anyone wants to take a look, please roast my code :)
Working on AI Coaching App to Make You Anti-Fragile, at https://www.reflectra.app/
Will work on improving the coaching agent to be more empathatic and context-aware. And user accounts.
just finished a organization project for my wife.
its a web app where you make boxes, add images or text of what's in the box. then get a qr code that you can tape to the box and scan to see the text or images in the web app.
hoping to make it a lot easier to look for things in the storage unit. instead of removing all the totes and looking in them. Just scan and see if the description fits what I'm looking for
A bunch of little electronic pin badges that I’m using to fund bigger projects
Currently in the works are a digital sand timer which can be used to track pomodoros (or any sequence of time intervals), and a Jovian orrery which displays the positions of Jupiter’s moons on a strip of addressable LEDs.
I have been working on a customizable AI voice dictation tool using Pipecat's framework to swap between many providers and models, including cloud or local.
Started off as an open source alternative to Wispr Flow for myself as I wanted to have more control over the formatting rules as well as model choice but after sharing with friends and presenting it at my local Claude Code meetup, I was encouraged to share it more widely.
The desktop app uses tauri so it is cross-platform compatible and I have tested it working on macOS and windows.
Building Contextify - a MacOS application that consumes Claude Code and Codex transcripts, stores them in a local sql db.
The main window uses Apple’s local LLM to summarize your conversation in realtime, with some swoopty UI like QUEUED state on Claude Code.
I’ve just added macOS Sequoia support and a really cool CLI with Claude Code skill allowing seamless integration of information from your conversational history into aI’s responses to questions about your development history.
The CLI interface contract was designed to mutual agreement between Claude code and codex with the goal of satisfying their preferences for RAG.
This new query feature and pre-Tahoe support should be out this week, but you can download the app now on the App Store or as a DMG.
I’m very excited about this App and I would love to get any feedback from people here on HN!
My Show HN: from this past week has a short demo video and a bit more info:
I'm looking to leverage the upcoming WebNN browser spec with my spreadsheet app. I think integration with vision, audio and language models opens a whole new world of possibilities as traditional spreadsheet apps lean more towards mainly numerical data.
A budget app called Ledga that shows your immediate past and forecast future for people living check to check. Desktop app with companion mobile web app.
site: https://ledga.us/
Subscription and Plaid features coming soon.
Free 60-day trial.
On Thursday I learned about ulid[0] which I think really neatly solves the problem of text representation for UUID v7. However, I also like the idea of prefixed ids, although I haven't used them in anger.
Yesterday I built most of a Postgres extension, using the excellent pgrx[1] project, that build on ulid to add prefixes. With it you get something like this
plid=# SELECT gen_plid('u');
gen_plid
---------------------------
u_06DHRQH6SJT7N2WEQK4910R
(1 row)
I haven't pushed it to GitHub yet, but it's fairly done at this point.
I was thinking about FizzBuzz and thought it might be cool to benchmark various LLMs to see the highest number they could go before they got it wrong. FizzBuzz is cool because you can test whether the model's can generalize to any other game (divisors of 3 and 7 instead of 3 and 5 for example).
Fun, short and sweet experiment to run over the weekend, with some mildly interesting results :)
Basically personalized meal planning and grocery integration. Since the Show HN I posted a couple months back I've been incorporating user feedback to add things like meal prepping, better ingredient reuse across meals, and cooking style preferences.
One of the biggest points of feedback has been adding more grocery stores but I'm really limited by who has APIs to actually integrate with, which is basically just Kroger and Instacart. Walmart has an API but ignored my API access request. Would love to hear if anyone has ideas on how to approach this.
We’re working on https://www.requoted.app/ — a simple quoting app for small trade businesses, especially fabricators, contractors, etc.
We kept seeing folks doing quotes in notes, spreadsheets, or texting numbers back and forth, then rebuilding everything again when something changes. ReQuoted is meant to make that part easier: build a quote from your phone, reuse materials/labor, send something clean, track, and revise without starting from scratch.
It’s intentionally lightweight — not an ERP, not trying to run the whole business.
My co-founder and I recently quit our jobs to focus on this full-time. Still early, but we’re already working with real shops and contractors and adjusting fast based on how they actually work.
Happy to hear feedback.
I'm building a temperature controlled dough proofing enclosure based on ESP32, an IKEA Samla storage box and an inexpensive silicone rubber heater. It's delightfully impractical-- purely a first hardware/MCU learning project and a holiday gift for a relative who I think will appreciate it.
I'm impressed by how far I can get "vibe making". Most of my professional experience is in high-level software, but AI gets me unstuck quickly when I don't know something specific to ESP-IDF or the hardware. As of today I've got a circuit tested, firmware nearly complete, and a custom PCB en route from JLCPCB.
One limitation I’ve noticed: ChatGPT struggles with the details of part selection (e.g. choosing specific temp/humidity sensors or connectors). Adding datasheets to the context helps a lot, which makes me wonder why this isn’t something the model can do or at least ask for.
An AI assistant plugin for Logseq. https://github.com/shovon/logseq-ai
It allows users to "chat" with their Logseq graph. Think of it like a "Cursor for Logseq". I hope people find it useful. I have on numerous occasion wished that I could have easily asked about a specific block on my graph, and would provide an intelligent response, also somewhat influenced by the contents of the entire graph. It's still a work in progress.
It's fully open source.
I built https://nofone.io . I ingest health insurance policies and provide insights to insurers on how to improve them and doctors to know what insurers expect to see in documentation and evidence. My hope is to improve the denial situation and standardize medical necessity criteria down the line.
It's essentially a book progress tracker. There are many apps that allow you to add the books which you are reading currently, but not at what pace. It's simple, no complicated stuff, no AI shenanigans.
Created as I was overwhelmed by the number of books I want to read and thought it would be helpful to plan ahead.
You add a book name, number of pages and how many pages you want to read in a day. It calculates and gives you the number of days and on which date you will finish. It's also flexible to increase the number of pages so that it can recalculate.
It's a PWA for now. Still working on notifications and stuff.
Published my first AUR package this week. It's called bleep, a simple interval timer that beeps.
The idea came from cooking bolognese. I needed something to remind me when to stir. So I wrote a small Go tool that just beeps at whatever interval(s) you set.
Then I kept adding stuff. Verbose mode with a live countdown, pause/resume with signals, and a JSON output mode that works with Waybar. That last one is actually my favorite part. I get a little timer in my status bar that changes color when it's counting, paused, or beeping. Click to pause. Works great for pomodoro or just keeping track of things while working.
I switched from Mac to Arch and wanted to try the whole AUR thing. Used GoReleaser to automate the build and publish. Took some fiddling but it works now.
https://github.com/Gioni06/bleep
AUR: yay -S bleep-bin
https://github.com/cdlewis/snowboardkids2-decomp
A matching decompilation of snowboard kids 2 for the n64. Why this game? Well it's awesome but also I wanted to work on a decomp project from scratch. I've written several blog posts about my experience for those interested. I hope to do more in the future, probably with less of an AI focus.
* Using Coding Agents to Decompile Nintendo 64 Games https://blog.chrislewis.au/using-coding-agents-to-decompile-...
* The Unexpected Effectiveness of One-Shot Decompilation with Claude https://blog.chrislewis.au/the-unexpected-effectiveness-of-o...
We are working on DB Pro, a modern desktop data workbench for developers and data engineers.
The focus is on going beyond a query editor and building a complete environment for working with data. Visual exploration, inline editing, dashboards, and Jupyter notebook style workbooks for queries, notes, and experiments all in one place.
We launched v1 a few weeks ago and the reaction has been genuinely jaw dropping. Downloads, feedback, feature requests, and some great long form discussions around real world data workflows.
We are documenting the entire journey through a public devlog series. The latest video covers the v1 launch.
Honestly, building a desktop app is so refreshing after spending a decade or so building web apps.
https://microyem.com/ canlı yem satış