← All workCASE STUDY · RIZM TECH · JEDDAH, SAUDI ARABIA
A WhatsApp sales agent that answers every lead, then hands it to a person.
Rizm Tech wanted inbound WhatsApp leads captured, qualified and followed up in one place, for its own sales and for its clients. I'm building the agent and the CRM behind it.
- ROLE
- AI engineer (contract)
- TIMELINE
- Sep 2026 – ongoing
- STACK
- WhatsApp Business API, Claude API, Next.js, PostgreSQL
- WAY OF WORKING
- 48-hour MVP sprints
THE BRIEF
Answer every lead, but only with what the business has approved.
Sales conversations in the Gulf happen on WhatsApp. The brief was to make sure every one of them is answered, qualified and tracked, without letting an AI say things the business hasn't signed off.
WHAT MADE IT HARD
- The agent can only answer from a brief the client has approved, and has to say so when a question goes beyond it.
- Saudi data-privacy law (PDPL) means consent and data erasure have to work from day one.
- Prompts change often, so a small wording change can't be allowed to quietly break a working conversation.
- When a person takes over, they need the whole conversation, not a summary.
WHAT I BUILT
- A WhatsApp Business API integration that receives and replies to messages in real time.
- An agent on the Claude API, grounded in the client-approved brief, that answers and qualifies each lead.
- A CRM in Next.js and PostgreSQL that captures, qualifies and follows up every lead in one pipeline.
- Handoff to a person with the full conversation attached.
- Automated test suites for both the bot and the CRM.
- Consent and data-erasure controls for PDPL.
HOW IT FITS TOGETHER
Lead on WhatsAppMessages at any hour
WhatsApp Business APIReceives and sends messages
AgentClaude API, answers only from the approved brief
CRM pipelineNext.js + PostgreSQL: capture, qualify, follow up
Sales teamTakes over with the full chat
Runs across every step: consent and data-erasure controls (PDPL)
Runs across every step: automated bot and CRM test suites
DECISIONS
DECISIONGrounded answers over open-ended chatWHYThe agent speaks for the brand to real buyersTRADE-OFFIt hands off more often instead of improvising
DECISIONTest the conversations, not just the codeWHYPrompt changes are frequent and easy to get wrongTRADE-OFFMore setup in the first sprint
DECISIONPrivacy controls from the first releaseWHYPDPL applies from the first messageTRADE-OFFA consent step before the conversation starts
DECISION48-hour MVP sprints with scope confirmed firstWHYThe partner sees working software fastTRADE-OFFRefinement runs alongside the next sprint
RESULTS
48hfrom agreed scope to a working MVP
1pipeline for capture, qualification and follow-up
Day 1PDPL consent step, live from the first release