Belumind Science
We build and run our own products, and we take on client projects in AI and software. Behind both is steady research into how software and AI are built and run — written up in full, including the questions we haven't answered yet.
The same skills behind our products and your project
Four areas of work, and none of them stops at a demo. Each one supports something already in use — one of our products or a client's system — or something we are working toward step by step.
Products & client work
We carry work all the way to something people use every day. Our own products — a PDF-to-Word tool for people handling financial reports, an assistant that answers from Vietnamese law, a Mac app that keeps everything on your computer — and custom AI and software for clients when nothing ready-made fits. Real use is where every design choice gets tested.
New technology
ARMT sends one file over every connection two devices have in common at once — from very slow to very fast — and when one connection drops, it simply shifts the work to the others instead of failing. Measuring how those connections really behave is the research behind it.
System design
We split systems into clear parts with agreed rules between them, and leave room to plug in pieces we have not built yet. Things going wrong is planned for as a normal event, not a surprise — which is mostly what decides whether a system still works in its second year.
AI that does the work
Reading scanned documents and rebuilding their tables exactly, answering questions straight from the original legal texts, and turning speech into text with translations and summaries. Our published sign-language model is just 0.74 MB and answers in 3.02 milliseconds on an ordinary processor — keeping AI small and checkable is a choice, not a limit.
What we ship
Bring us a problem
Alongside our own products, we take on client projects in AI and software — problems no ready-made tool solves, where AI has to run somewhere unusual, or where a system has to keep working even when part of it breaks. Here is some of what that has looked like.
Sustainability (ESG) reporting arrives as a mountain of documents — reports, certificates and forms in whatever format each supplier sends. We built the system that turns that pile into organised data anyone can check: it pulls out the figures that matter, keeps the tables and context they came from, and links every number back to the exact page it was read from.
Opening an account online only works if the bank can be sure who is on the other side of the camera. We built the AI that makes that call: it reads Vietnamese ID cards, checks that the face in front of the camera matches the photo on the card, and tells a real, present person apart from a photo, a recorded video or a computer-made face. It runs on whatever phone the customer already owns, in whatever lighting they happen to be in, and it has to be right — the cost of a wrong answer is a fraudulent account or a customer turned away.
Measurements stop being accurate when the real signal is buried in noise. We built the AI that finds it — separating the real signal from interference so the measurements that follow can be trusted. Cleaner input, fewer wrong readings, and tuned for the real conditions it works in rather than a tidy lab test.
Research and engineering, in the open
Papers with the numbers to back them up, and articles that describe a problem before we have the answer. Both show how we think — they are not advertising for our products.
Latest paper
A way to translate from any spoken language straight into sign language, by first capturing what a sentence means rather than its exact words. It picks out the key details correctly 99.03% of the time across four languages, using a tiny 0.74 MB AI model that answers in 3.02 milliseconds on an ordinary processor.
Latest write-up
12 min readActivation sparsity, a Metal kernel, and what did not work along the way. On ProSparse-Llama-2-7B only 11.2% of FFN neurons fire per token; skipping the rest costs +0.44% perplexity and makes the FFN about 2.05× faster on an M4 Pro. SwiGLU models, fixed clusters and cheap predictors did not pay off.
The products, the papers, and the engineering write-ups all point at each other. Pick whichever one you would rather read first — or just get in touch.
or email hello@belumind.com
