Sarvada Labs: WhatsApp Commerce Agent
Concept build, not a client project
Built to prove out the pattern before pitching it to D2C clients — a WhatsApp agent that takes a vague product query ("something light for summer, under $40"), asks a couple of clarifying questions, recommends from a catalog, and hands off to a human only when needed.
This is a self-funded, self-built concept demo — not a client engagement, and no client data was used. It exists to prove the pattern before we scope it as a paid pilot for a D2C client.
Where it started.
Before we ask a D2C founder to pay for a commerce agent, we think we should be able to show one running. So we built the pattern ourselves, end to end, on our own time.
The hard part of conversational commerce isn’t the chat — it’s the moment a customer types something vague like “something light for summer, under $40” and expects a useful answer. Handling that well takes catalog structure, a couple of smart clarifying questions, and knowing when to stop guessing.
No client asked for this, and no client data touched it. That’s the point: we wanted the failure modes to happen to us before they could ever happen to a customer’s store.
What we did, step by step.
Design for the vaguest possible input
The agent is built around under-specified queries. It extracts whatever constraints it can — occasion, budget, scent notes, skin type — and asks at most two clarifying questions before it recommends anything.
Catalog matching, not free generation
Recommendations come from a structured catalog: deterministic filters narrow the field, then the model reranks what survives. The agent cannot invent a product or a price — if it's not in the catalog, it doesn't get suggested.
Human handoff as a designed feature
When confidence drops — an unusual request, a complaint, a wholesale inquiry — the agent stops and hands the conversation to a human with the transcript and its current shortlist attached. Escalation is part of the design, not a failure state.
Evaluated like client work
Same QA protocol we'd run for a paying client: an evaluation suite of vague, edge-case, and adversarial queries, deterministic fallbacks, and regression runs on every change. If we wouldn't ship it to a client, it doesn't count as proven.
The vague-query test.
Watch the agent handle a real vague query end to end — from "something light for summer" to a checkout-ready recommendation, with clarifying questions in between.
Sarvada Labs Agent
online · concept demo
Demo video coming soon
Working concept demo on a test WhatsApp Business number. It serves as the reference implementation behind the commerce-agent pilots we scope for D2C clients — self-funded, self-built, and clearly not a client outcome.
Want this pattern built for your catalog?
This concept demo is the reference implementation — a client pilot is scoped around your actual products and channels.