13 community apps · carebun.com
Small apps, made with care, for the people around us.
Free tools for learning something properly, finding your people, and getting through the paperwork of a new country. Each one has its own address on carebun.com, and the parts underneath are open.
- interactive.carebun.com50 ms
- rag.carebun.com222 ms
- vectors.carebun.com47 ms
- memory.carebun.com784 ms
- courses.carebun.com472 ms
- bigindianhub.com785 ms
- checkin.carebun.com1399 ms
The apps
Every app, in one place.
Learn
Read, understand, remember.
5- 50 ms Carebun Reading Research, writing and books in one place, with shelves and recall. Everything to read on carebun: the research with its articles, experiments and papers, the essays, and the books, each with a map, a cheat sheet and cards. Reading is open to everyone. Sign in to keep a shelf, take the quiz, and write what you remember; recall comes back, spaced, every day. research, writing, booksshelves and spaced recallopen without an account live at interactive.carebun.com 30 days
- 222 ms RAG, the complete book A hands-on RAG course where every number is real. Retrieval-augmented generation from the ground up: chunking, embeddings, retrieval, reranking, evaluation, economics and speed, and the 2026 shape of the whole thing. hands-onreal measurementsfree to read live at rag.carebun.com 30 days
- 47 ms Thinking in Vectors From “what is a vector?” to retrieval over a hundred million chunks. A book written as a series of posts. Every idea is built from something you already understand, explained with an everyday example, then pushed until it holds up in production. 66 chapters10–15 min eachwith a map and cards live at vectors.carebun.com 30 days
- 784 ms Memory for AI Agents How an AI assistant remembers you, in plain English. A card file with two dates, a secretary who never tears a page out, and a librarian with three ways to search. The plan, and eleven short reads. 13 pieces~49 min in totalno jargon live at memory.carebun.com 30 days
- 472 ms Scaling Laws & Agent Memory Two ground-up courses. Why bigger models get better, and by how much; and how agents keep what matters. Built step by step, for people who want the reasoning and not just the result. 2 coursesground-up live at courses.carebun.com 30 days
Belong
Find your people, show up together.
2- 785 ms bigIndianHub Indian Kansas City, all in one place. Desi restaurants and groceries, temples, garba nights and Durga Puja, doctors, realtors, groups, sports and a kids’ corner: a community platform for the Indian diaspora in Kansas City. events & placesKansas Citycommunity-run live at bigindianhub.com 30 days
- 1399 ms Event Scanner One QR code at the door. Check-in for any event: guests register and get a QR code, the organiser prints a sheet or scans a badge, and everyone knows who came. QR check-infor any event live at checkin.carebun.com 30 days
Navigate
Paperwork and systems, made legible.
1Build in the open
The parts underneath, shared.
5- 784 ms HanuDB A document database with search that runs for free, and cannot lose your data. The database under the newer apps here. JSON documents, filters, sorting, full-text search and aggregations on the smallest VM Google gives away; the truth lives in a storage bucket, so the machine can be deleted at any moment and a new one rebuilds itself. Open source. under 1 ms~1¢ a month34,800 writes/s live at db.carebun.com 30 days
- 49 ms Research An article, the experiments, and a paper. Every research piece on one page, each with an article to read first and a paper. Open work shows everything: the experiments, the ones that failed among them, and the code. Closed work shows its results. article, then paperopen work shows its experiments live at lab.carebun.com 30 days
- 66 ms hanumemAI Long-term memory for AI assistants: remember everything, inject almost nothing. A memory layer measured on three public benchmarks, every question, with every model and token count named. The highest published score on LoCoMo, 96.1, reading 1,533 context tokens per question. Proprietary: the results and a paper are public, the ideas are taught in a book, and the library is on request. LoCoMo 96.11,533 tokens per questionon request live at carebun.com 30 days Read the write-up: how it works and what it looks like
- 784 ms Gobiman See a Postman collection as a map, run it, and search every response. Turns any Postman collection into an interactive mind map and a client-to-API sequence diagram, runs the requests in order, and opens every response in a collapsible, searchable JSON viewer. Try it here with a demo or your own collection, nothing stored; or clone it: one HTML file and one Node file, no install, no account. mind map + sequence diagramsearches inside responsestry it, or clone it live at gobiman.carebun.com 30 days Read the write-up: how it works and what it looks like
- 785 ms gptimg A command that asks the ChatGPT app to draw something, and saves the file. One Swift file for the Mac. Type gptimg “a fox reading a newspaper” and it drives the ChatGPT desktop app the way a screen reader would: finds the prompt box, sends the words, waits for the picture, opens it, presses Download and saves it where you said. No API key, no browser automation, nothing to install but the app itself. Open source. no API keyone Swift fileprints the saved path live at gptimg.carebun.com 30 days
Nothing matches that.
The database underneath · HanuDB
The bucket is the database. The machine is only a cache.
A document database with search that runs for free, and cannot lose your data. JSON documents, filters, sorting, full-text search and aggregations, on the smallest virtual machine Google gives away. The truth lives in a cloud storage bucket, so the machine can be deleted at any moment and a new one rebuilds itself.
The free VM
- Fast: reads and writes in well under a millisecond
- Fragile: one machine, one disk
- Small: 1 GB of memory, a quarter of a core
- Free
Serves every read and every write.
The storage bucket
- Slow: 40–70 ms per request
- Indestructible: replicated, optionally across regions
- Huge: as big as we like
- Cents per GB per month
Holds the hash-chained log and snapshots: the truth.
Think of it like writing in a notebook and photographing each page as you go. You read and write in the notebook, because that is fast. If the notebook falls in a river, you print the photos.
- Every write goes to a log file first. The write-ahead log: append-only, in order, the one thing everything else is rebuilt from.
- One disk sync covers many writes. Group commit, the elevator trick: 32 clients make 34 800 fsynced writes a second.
- Two engines read that one log. An LSM tree for get-by-id; an inverted index for filters, sort, search and aggregations.
- Once a second, the log is shipped to the bucket. Compressed, hash-chained segments; idle costs nothing.
- Snapshots keep recovery short. A rebuild replays only the log after the newest one: 4.5 GB in 40 seconds.
- Only one writer, guaranteed by the bucket. “Only if it does not exist” uploads; no clocks needed for safety.
- It fits in one gigabyte. Small caches, a memory guard, and streaming everywhere.
- writes per second
- 34 800
- reads per second
- 46 200
- to rebuild from 4.5 GB
- 40 s
- bucket cost, a day of heavy testing
- ~1¢
Measured on a container shaped like the free VM (one core, 768 MB), with a real disk sync on every write and a wipe-and-rebuild from the bucket at the end of every run.
curl -fsSL https://raw.githubusercontent.com/madhudream/hanudb/main/install.sh | sh
hanudb serve --bucket mem:// --dir ./data --listen 127.0.0.1:8080
How we build
Made the way a neighbour would make it.
Free to use
If it helps people, it should not need a card. Community apps here are free, and stay that way.
Small on purpose
Tiny compute, few moving parts. The newer apps share one small database on a free-tier machine, so they are cheap to keep running for years.
Open where we can
The database is open source; the experiments are published with their numbers, including the ones that did not work.
Private by default
No ads, no tracking scripts. What you write, like your recall notes in Carebun Reading, belongs to your account and no one else.
Suggest one
What would help your community?
A sign-up sheet for the temple kitchen, a way for new arrivals to find a doctor who speaks their language, a reading group that actually remembers what it read. Tell us; every note is read.