Jeweller Clienteling Agent: OpenClaw Runtime Proof

One-Line

Jeweller Clienteling Agent is an OpenClaw-backed concierge demo for jewellery enquiries that turns WhatsApp/Instagram-style messages into safe customer replies and staff briefs.

Why It Matters

Jewellery retail has a high-trust front-desk problem: customers ask about bespoke pieces, stock, repairs, order status, sourcing, complaints, appointments, and prices through informal channels. A generic chatbot is risky because it can promise prices, stock, delivery, authenticity, or repair outcomes without approval.

The useful system is a clienteling assistant: fast enough to reduce missed enquiries, structured enough to prepare staff, and careful enough to escalate commercial or high-risk promises.

What I Built

• A deterministic enquiry engine for intent detection, missing-field collection, authority level, handoff rules, customer replies, and staff briefs.
• Safe handling for bespoke, stock, sourcing, repair, care, complaint, order-status, and price-band scenarios.
• Fixture-backed checks for stock, price bands, order-status dry-runs, and repair-intake guidance.
• Dry-run WhatsApp and Instagram channel payload contracts with send_enabled: false.
• A local OpenClaw runtime adapter that wraps the engine and proves the gateway path invokes the real repo code.
• Eval scripts that compare local gateway output with deterministic engine output instead of trusting prompt-shaped responses.

Deployment Angle

This is useful for AI Deployment / FDE roles because it shows the same deployment instinct outside education:

• start with the messy workflow where users already message the business;
• define which decisions the agent may make and which require approval;
• turn unstructured messages into structured operational briefs;
• prove the runtime path, not only the static prompt;
• keep live channels and customer data behind explicit approval.

Technical Shape

• Interface model: WhatsApp/Instagram-style inbound enquiries.
• Runtime: local OpenClaw gateway plus repo runtime adapter.
• Core engine: Python deterministic policy and response generation.
• Proof: deterministic tests, golden scenario evals, dry-run channel contract tests, and local gateway runtime evaluation.
• Guardrails: no exact price promises, no stock holds, no repair outcomes, no sourcing guarantees, no private supplier leakage, and staff approval for commercial commitments.

Current Proof Level

The current proof level is runtime-proven-demo: local OpenClaw invokes the repo runtime adapter, the adapter invokes the deterministic engine, and runtime output is checked against deterministic expectations. It is not a live WhatsApp/Instagram deployment and does not use real customer data.

Best Interview Angle

This project is a good example of deployment judgement. The hard part is not making a jewellery chatbot sound polite; it is deciding what the system is allowed to do, proving the runtime path, and keeping commercial promises behind the right approval boundary.