Fathin Dosunmu
Co-founder
The problem
Small teams rarely lack information. They lose the thread between an inbox, a Slack conversation, a meeting note, and the document someone updated two weeks ago. The cost appears later: a missed promise, a repeated decision, a client risk nobody connected in time.
The product needed to remember across those tools without becoming another place people had to maintain.
The line we would not cross
AI handles the mechanical parts brilliantly and the judgment parts badly. Judgment is the entire thing people remain responsible for.
The easy version would have been a chatbot that gave confident answers and quietly took action. The stronger product separates mechanics from responsibility: ELAV can retrieve, connect, cite, draft, and propose. A person approves consequential changes and actions.
What I worked on
I worked across the product shape and the engineering stack, from deciding how company memory should behave to making source connections, evidence, permissions, approval paths, and production delivery real.
- Context with receiptsAnswers tied back to their sources, with provenance people can inspect.
- Memory with permissionsCompany context kept tenant-scoped and aligned with source access.
- Action with approvalDrafts and proposals that keep consequential decisions with a person.
- Production over demosSource health, ingestion recovery, identity, onboarding, and operational seams.
Trust was architecture, not a footer claim
A useful company memory touches sensitive conversations and decisions. That made security part of the product model: permission-aware retrieval, tenant boundaries, approval gates, visible sources, and explicit handling of external input rather than blind trust in whatever entered the system.
The interface could stay simple because those constraints lived underneath it.
The result
ELAV is built for teams of 5 to 50 where company context no longer fits in everyone’s heads. It can show what changed, where an answer came from, and what is waiting for a decision without pretending the AI should make that decision itself.