Judge the engineering, not the adjectives.

The problem

A multi-branch maintenance contractor producing dozens of quotations and bills of quantities each month by hand — re-keying the same rates, scope items, and terms in Word and Excel on every job, with pricing that drifted between estimators.

The system we built

An ingestion step reads the incoming scope or RFQ; a structured store holds approved rates and reusable line items; a generation step assembles the quotation and BOQ from approved templates against current rates; and a human review gate signs off before anything leaves the building. Every output is versioned and traceable back to the source request.

Engineering decisions

Generation is constrained to approved rate tables rather than free-form model output, so numbers cannot hallucinate or drift. The human approval step stays in the loop because pricing carries commercial risk — the system removes the typing, not the judgement.

The problem

A beverage distributor moving cash and stock through a field delivery team, where custody, stock reconciliation, and proof of delivery were tracked on paper and trust — leaving room for leakage and disputes that only surfaced weeks later.

The system we built

A delivery-confirmation flow captures a per-delivery confirmation (OTP / signed handover) at the point of exchange; a reconciliation layer matches dispatched stock against confirmed deliveries and cash collected; and exception reporting flags mismatches for a supervisor the same day. Every handover is logged and attributable to a person and a time.

Engineering decisions

The confirmation step was built to be hard to bypass — a control only works if it can't be skipped. Mismatches surface as exceptions rather than blocking the route, so the field team keeps moving while the office gets visibility.

The problem

An appointment-based business — clinic, salon, or studio — paying for leads across web forms, Instagram, and portals, then losing them to slow or after-hours replies from a team that can't watch the inbox around the clock.

The system we built

A single WhatsApp entry point answers instantly, runs a qualification script (service, timing, budget, location), books into the calendar, and logs the lead to the CRM or a live sheet — with automatic follow-up on leads that go quiet. English and Arabic handling sits at the model layer, with human handoff on anything the script can't confidently resolve.

Engineering decisions

Instant acknowledgement first, qualification second — speed-to-first-reply is what actually saves a paid lead. The bot escalates to a human rather than guessing when intent is ambiguous, so edge cases don't turn into bad bookings.