A reusable AI/ML capability that products can adopt when the use case and safety boundary fit. Kasekor inside Chomkar is the current verified implementation.
Most teams rebuild the same AI plumbing for every product. We refuse to.
Prompts, voice, routing, guardrails, and evals are reusable building blocks. We verify one narrow capability first, then make it available where it genuinely fits.
A product may reuse a verified component instead of starting from zero. Products without a justified AI need remain independent of the layer.
The advantage is disciplined reuse: each proven component can lower later delivery cost without turning a roadmap assumption into a shipped claim.
The first thing running on the layer is already in a real product's hands.
Kasekor is a Khmer lesson-helper living inside Chomkar — it explains farming concepts to growers in their own language. It is firewalled from anything commercial: it never touches price, demand, or quantity. Teaching is its only job, and that boundary is deliberate.
The layer is run like a product, not a help-desk. Hongleng is its lead — with the right to say "not this week."
A product team brings a real need — a Khmer explainer, a classifier, a voice flow — and writes down what success looks like.
Requests are weighed against everything else the layer serves. Some ship now; some wait. "Not this week" is a real and frequent answer.
What gets built is designed to serve more than one product — a shared prompt, model route, or guardrail — not a one-off bolted onto a single app.
The component lands in the layer, the requesting product consumes it, and the next product inherits it for free. The roadmap moves on.
One owner, accountable for the reusable capability and every approved adoption.
Leads the reusable AI capability, verifies its boundaries, and decides what is ready for another product to adopt.
The layer is only as real as the products on it. Start with the one it lives in today.