An AI-powered automation engine that takes messy inbound work — WhatsApp messages, emails, form submissions, support tickets — classifies what it is, calls the tools it needs, and routes the outcome to the right place. Humans approve the edge cases, not the routine.
Orders, support questions, delivery issues and partner enquiries all landed in the same WhatsApp inbox and shared email. Someone had to read every message, work out what it was, and copy the details into the right system — several hundred times a week.
Each recipe is a versioned workflow in Rafiki — a trigger, some AI reasoning, optional branching, and one or more tool calls. They can be edited without redeploying the app.
Reads WhatsApp and email, classifies intent and urgency, answers common questions from a knowledge base, and escalates the rest with full context.
When an order is paid, sends a confirmation, books delivery, creates the ops task, and follows up automatically if the order stalls.
Reads supplier invoices from email attachments, extracts line items and totals, validates against the PO, and posts to the accounting system.
Scores inbound leads against fit criteria, writes the summary to the CRM, and books a call for the ones worth a human's time.
Every morning at 07:00, pulls yesterday's numbers, writes a plain-language summary, and posts it to the team channel.
When someone asks a policy or contract question, retrieves the answer with citations — and hands off to a human if confidence is low.
No black boxes. Each step in a workflow logs its input, its output and its decision, so when something behaves unexpectedly you can see exactly why — and replay it after a fix.
WhatsApp, email, webhooks, forms, cron schedules and database events all start the same kind of workflow.
Classification, extraction and summarisation return typed, validated JSON — not free text you have to parse later.
High-confidence runs proceed automatically. Anything uncertain goes to a human queue with the reasoning attached.
Built-in connectors for M-Pesa, WhatsApp, Slack, email, CRMs, Sheets and your own internal APIs via a typed tool registry.
High-stakes actions — refunds, large orders, outbound comms — can require a named human to approve before executing.
Every run is stored with its full trace. Fix a workflow, replay the failed runs, and confirm the change did what you expected.
“It doesn't replace the team — it just stops them doing the same three things four hundred times a week.”
— Operations lead, services deploymentI build AI automations that handle the routine, escalate the edge cases, and log everything — so your team stops being a message queue.