Earendel is a product studio for companies betting on AI. Most agencies will build you a demo. We embed with your team for 8–16 weeks and leave you with a working product, the evals to trust it, and the engineers who understand every line.
Strategy, design, engineering — under one roof, on one timeline.
We're opinionated about scope. Every engagement pairs a designer, an engineer, and a strategist. No layer-cake handoffs, no unaccountable "discovery" months.
Where AI actually earns its keep in your product — and where it doesn't. Ends with a prioritized, costed plan.
2-week sprint · fixed fee 02Search your company actually uses. Retrieval, ranking, citations, evals — the boring parts done right.
6–10 weeks · production 03Workflow agents with guardrails you can audit. Tools, memory, eval harnesses, human-in-the-loop.
8–12 weeks · embedded 04Interfaces for non-determinism. Streaming, uncertainty, undo, trust. We've shipped more than we've designed.
design lead embedded 05Pipelines, vector stores, feature stores, observability. The plumbing that makes AI reliable in production.
4–8 weeks · handover 06One-day intensives for product and engineering teams. No slides about the future — working on your stack.
on-site or remote 07Senior AI engineers and designers inside your squad, 3–6 months. Same calendar, same standup, same Slack.
from $14k / mo · per seat 08 · our productAn AI website platform for businesses that need a site, not a CMS. See it below.
Visit sites.earendel.et →A small selection. Some clients we can name, some we can't — most of this is in production revenue.
Compliance-aware search over 14 years of policy docs. Bilingual, cited, with red-team evals run weekly.
↗ 2026$2.1M/yr of analyst hours recovered. Every action is reviewable, reversible, logged.
↗ 2025Our own flagship. 400+ restaurants, clinics and shops live. Our playground for AI UX.
↗ 2025First-response time down 68%. A tiny model doing the boring work, not the expensive one.
↗ 20252,400 labeled conversations, 60 adversarial suites, weekly regression dashboard.
↗ 2024The tokenizer most Ethiopian AI companies are now quietly using. Apache 2.0.
↗We dogfood every opinion we have about AI UX. Our flagship:
Tell us about your restaurant, clinic, or shop. We research, write, photograph, design, and ship a proper website in 48 hours — then you edit it in the portal we built. No templates. No drag-and-drop. No "AI website builder" slop.
An open eval workbench for product teams. Write cases in YAML, run nightly, alert on regressions.
Invite-only · Q3 2026A research track. Public checkpoints when we're proud of them. No AGI takes.
Updates on our writingEvery engagement follows the same rhythm. You always know what's being built, by whom, and why.
Two weeks. We read your code, talk to your users, and write a short memo. You keep the memo whether we continue or not.
Week 0–2A working thing on your data, in your stack, behind a feature flag. Not a Figma. A deploy link.
Week 2–5Evals, guardrails, observability, docs, runbook. The unsexy work that separates a demo from a product.
Week 5–10Your engineers own it. We run a 2-week shadow period, then we're on call for 90 days at no charge.
Week 10–12We're four partners and a rotating bench. If you hire us, partners do the work — not account managers.
Builds the hard parts. Previously staff eng at Andela. Spends weekends on slow-cooked doro and faster React.
Makes interfaces for non-deterministic systems. Wrote our house style. Doesn't believe in drag-and-drop.
Runs the research track. Published on low-resource NLP. Has opinions about tokenizers that will outlast this decade.
Keeps engagements on the rails. Spent a decade in healthcare ops. Will ask the question nobody wants to ask.
Plus 8 engineers, designers, and researchers who rotate through the bench. We hire slowly. If that sounds interesting, work@earendel.et.
What we publish when we have something actually worth saying. Monthly at most.
Most "AI products" are demos with a Stripe link. A real product has evals, runbooks, and a plan for when the model is wrong — which is always.
Read the essay →A short, honest description of the problem is enough. We'll reply within two business days with a proposed call, or tell you straight that we're not the right team.