Client work · Lebanon

A CRM that speaks any language

TopCrete International manages hundreds of WhatsApp conversations a day: orders, leads, site inquiries, voice notes. We built them a private AI platform that reads all of it, understands the Lebanese dialect, and turns every conversation into a priced, mapped, tracked deal.

TopCrete International logo

Built for TopCrete International

Lebanese concrete & construction materials, operating across every region.

It listens.

The platform ingests WhatsApp .txt exports in both iOS and Android formats, including chats written in Arabizi (Arabic rendered in Latin letters with numerals like 7 for ح and 3 for ع, the default way most Lebanese text each other). Voice notes are transcribed via Gemini across regional accents (Beirut, Mount Lebanon, the North, the South, the Bekaa) and fed into the same analyzer that reads the text messages. Contact, project, area, thickness, coating: all extracted, all structured.

It understands the math.

From a voice note saying “bḋna 300 mitr fi Baabda, 12 santim, ma3 coating”, the system extracts 300 m², 12 cm thickness, Baabda location, and a coating upgrade, and returns a priced quote using the client’s five-tier area pricing (50 to 100, 100 to 200, 200 to 500, 500 to 1000, 1000+ m²), thickness margins for 8 / 10 / 12 cm, per-m² coating, and optional pump. No forms. No re-keying. The sales team reads their WhatsApp; the platform does the math.

It runs the pipeline.

Every analyzed chat lands on a map of Lebanon, pinned by region. Deals move through a Kanban board grouped per location: Prospect, Qualified, Negotiation, Closed. A follow-up queue surfaces action items before they go cold; a risk-flags view catches silent leads, stalled negotiations, and overdue payments. An analytics dashboard rolls up KPIs and pipeline projections, with an AI advisor that summarizes the numbers and flags what to look at first. What was buried in a WhatsApp scroll is now a dashboard.

How it’s built

  • Python 3.11 + Streamlit

    11-page operator UI, no build step, fast iteration.

  • SQLite with WAL

    Single source of truth. Auto-migrations, cascade deletes.

  • Gemini 2.5 Flash

    Chat analysis and voice-note transcription in one model.

  • Ollama + NVIDIA NIM

    Local-first AI advisor with a cloud fallback path.

TopCrete couldn’t use an off-the-shelf CRM. No SaaS speaks Lebanese dialect. No template pricing tool handles concrete by square meter across five area tiers with thickness margins. This is what we build: custom AI systems for businesses that the generic tools don’t fit.

Want something purpose-built for your business?

Talk to us →