A small selection. Some clients we can name, some we can't — most of this is in production revenue, not a pitch deck.
A retrieval-grounded copilot for 4,200 branch staff.
Compliance-aware search over 14 years of policy documents. Bilingual (Amharic/English), every answer cited, red-team evals run weekly against a growing adversarial suite. Branch staff resolve policy questions in seconds instead of escalating to head office.
Invoice reconciliation agent, 8 languages.
An agent that matches, flags, and files invoices across 8 languages and 3 ERPs. Every action is reviewable, reversible, and logged — human-in-the-loop by default, autonomy earned per workflow.
A website generator for real businesses.
Our own flagship. 400+ restaurants, clinics and shops live in production. Research, copy, design, and deploy in 48 hours — then owners edit in a portal we built. Our playground for AI UX.
Claims triage across 14,000 monthly cases.
First-response time down 68%. A tiny classifier doing the boring work, not the expensive one — with a clear escalation path for everything it is not sure about.
Eval harness & guardrails for a user-facing chat.
2,400 labeled conversations, 60 adversarial suites, a weekly regression dashboard the whole team reads. The chat ships weekly because the evals catch what reviewers miss.
Amharic tokenizer & instruction-tuned release.
The tokenizer most Ethiopian AI companies are now quietly using. Apache 2.0, maintained, documented. Our research track pays rent in the open.