Skip to content
Achraf Abderrazik
ENFR
All work
Real-world · InternshipCase study 01 / 05

Safe Invest AI Lead Automation

Multilingual LLM lead pipeline

Messages prospects send on WhatsApp, Instagram or Messenger, in French, English or Darija, are turned into structured CRM leads automatically.

Organisation
Safe Invest Property · Marrakech
Role
AI Engineering Intern · Sole developer
Timeline
July – August 2026 · 9 weeks
Languages
French · English · Darija

Built during an internship. The agency’s code, data and conversations are confidential and not shown here.

Extraction accuracy
98%
How often the extracted lead data matched the annotated reference, measured on 96 messages in French, English and Darija.
Fewer LLM calls
≈4×
The drop in LLM call volume once messages sent in quick succession were grouped before extraction.
End-to-end scenarios
13/14
Full runs from incoming message to CRM lead, validated after voice-note transcription and lead↔property matching were added.
CRM fields per lead
9
The fields in each CRM record built from a conversation. 3 of them are computed in code rather than by the model.

01Context

Safe Invest Property is a real-estate agency in Marrakech whose prospects write in through WhatsApp, Instagram and Messenger, in French, English and Moroccan Darija. During a 9-week internship I was the sole developer of the system that turns those conversations into CRM leads.

02Problem

Inquiries arrive as free-form conversations: several short messages in a row, voice notes, three languages. Each one has to become a complete, consistent lead record, without creating duplicates when a platform delivers the same webhook twice.

03Architecture

  1. 01Channels

    WhatsApp · Instagram · Messenger

    Inbound webhooks

  2. 02Intake

    Idempotent ingestion

    Duplicate webhooks absorbed by DB constraints

    02Buffer

    Burst debouncing

    One LLM call per burst

  3. 03Voice

    Voice-note transcription

    Audio to text

  4. 04LLM

    Structured extraction

    Claude API · forced tool use

    98% accuracy

    04Code

    Deterministic fields

    Computed, not generated

  5. 05Data

    PostgreSQL

    5-table schema

  6. 06Match

    Lead ↔ property matching

    06Output

    Structured CRM lead

From three messaging channels to a structured, matched CRM lead.

04My contribution

Sole developer, end to end: pipeline architecture, the PostgreSQL data model, LLM extraction and its evaluation, voice-note transcription, the lead-to-property matching engine and the production deployment.

05Technical decisions

  • 01

    Structured output, not free text

    The LLM returns each lead through a forced tool call with a fixed schema, so every response maps straight onto CRM fields instead of being parsed out of prose.

  • 02

    Deterministic logic stays in code

    Fields that follow fixed rules are computed in code rather than generated by the model, which removes a whole class of errors by design.

  • 03

    Idempotency in the database

    Constraints in the 5-table PostgreSQL schema absorb duplicate webhook deliveries, so a repeated delivery doesn’t create a second lead.

  • 04

    Debounce message bursts

    Prospects often send several short messages in a row. They are grouped before extraction, so a burst costs one LLM call instead of one per message.

06Results

  • Deployed to production: conversations from three messaging channels become CRM leads automatically
  • One pipeline for French, English and Moroccan Darija
  • Voice notes transcribed and leads matched to the agency’s properties
  • Duplicate webhook deliveries absorbed in production by idempotency constraints

07Stack

  • Claude API
  • Structured outputs
  • PostgreSQL
  • Webhooks
  • WhatsApp · Instagram · Messenger

08Source code

The source code isn’t public: it was written during an internship and stays private. Other projects are on GitHub.

Next case studyAMANToxic speech detection in Moroccan Darija

Building something similar?

Start a project