The answer is already in your documents
Most of what an institution is asked every day is written down somewhere already, in a handbook, a policy, a timetable or a past paper. It is just not findable at the moment somebody needs it. Draimo AI, built by Draimo Inc, turns that material into an assistant that answers from it and shows its working.
It started at Strathmore
Draimo began as MariaChat, an assistant for one university: Strathmore, in Nairobi. The problem there was never that the answers did not exist. Admissions requirements, fee structures, exam regulations, timetables and term dates were all written down already. Getting one of them meant joining a queue, finding the right office, or sending an email and waiting.
So the first rule written into Maria is still the one that defines her, and it is enforced in the system prompt to this day: never send somebody to a website, an email address or an office. Answer from the material. Where the material does not cover the question, say exactly that and keep helping, rather than handing the person back the queue they came to avoid.
Everything since is that same assistant, generalised. It answers in nine languages, turns course material into exam preparation, writes documents your staff can send, and researches questions with sources you can open. One platform, scoped to one institution’s own material at a time.
How it is built
Answers come from your documents, with citations
Every response is retrieved from your own material and cites what it used, so a reader can check it. Where the corpus does not cover a question, the assistant says so and points to who does, rather than filling the gap with something plausible.
One corpus, one embedding model, never mixed
Vectors from two different models are not comparable, and ranking across them fails quietly rather than loudly. Every document records which model wrote it, and retrieval refuses to rank across them. It is the kind of discipline nobody notices until it is missing.
Your users' language, not a translation layer
Nine languages, each a full input and output language rather than a translated shell. A corpus that is not in English is retrieved under a model trained for that, because an English model scores a correct non-English answer barely above an unrelated one, and a margin that small does not rank.
Priced in the currency your users are charged
Kenyan Shillings through Paystack across Africa, local currency through Stripe everywhere else, with usage metered per request and reconciled against what was actually spent. Billing is not an afterthought bolted to a demo.
Data locations, provider terms and model-use practices vary by feature and are set out in our legal and compliance materials.
Bring it to your institution
Tell us what your people ask most and what you have already written down. We will show you what an assistant grounded in it would answer.