AI functions in production:
AI agents on WhatsApp.
We do not wire a chatbot to a vendor's API. We build the AI function inside your app: an orchestration you own, documents retrieved rather than everything sent, an output checked against a schema, a fallback to a human, an evaluation set written with you and a cost per request read every week. In French, English, Arabic and darija, hosted where your contracts require.
See how we build.
Six AI functions, and how each one is built.
Click a function to see what the user gets, how it is built, and with which models. Each function has its “how we build” guide on the blog.
AI agents & RAG
Agents on WhatsApp, in the back office or inside the app, with tools (order, invoice, appointment), written limits and a measured handover to a human. RAG on your private documents: chunking, index, retrieval of only the useful passages, an answer constrained to cite or to say “I don’t know”. Versioned instructions, output schemas and a full log, in your repository, not in a vendor’s console.
- WhatsApp Business agent: order, tracking, complaint, with the 24-hour window, approved templates and cost per message
- Search in contracts, procedures and product sheets, with the source cited; “I don’t know” when nothing is found
- Sorting incoming messages into closed categories, human queue for the uncertain
- Internal assistant on the ERP and accounting, read-only tools first, action after validation
- Generation of quotes, replies and file summaries reviewed before sending
- Handover to a human below a confidence threshold, with context, measured takeover time
What can be verified, not what is promised.
A “how we build” guide per function
AI agent on WhatsApp, AI in the app, search in Arabic and darija, OTP login, offline mode, hosting, delivery, integration: eight published guides with architecture, figures and calendar. Read how we do it before calling.
Clutch 5.0, two verified reviews
Clutch interviews every client before publishing a review. Profile linked in the footer; ask us for the two clients.
NOOR, an AI platform of our own
Our sovereign AI platform runs on the client’s servers, offline if needed, in Arabic, French and English. Public product page, public skill card.
The code, models and keys are yours
Repository, vendor keys, accounts and domains in your name from the first delivery; a demo every two weeks on an environment you open yourself.
Twelve weeks for two or three AI functions. The first without code.
Each function has a known input, a constrained output, a threshold below which a human takes over, and an evaluation set written with you before the first line of code.
Functions, limits and evaluation set
Weeks 0 to 2The two or three functions chosen, what the AI does, what it never does, the fallback threshold, where the data lives; signed. Then sixty real cases with the expected answer, written with your manager: the acceptance test for everything after.
Build and pilot
Weeks 2 to 9Orchestration and the first function, the one that replaces the most re-entry, then the next ones on the same layer. Pilot with real users and active human fallback; share handed to a human, cost per request and latency read every day. A demo every two weeks.
Tuning, supervision, operations
Weeks 9 to 12, then ongoingCost goes down without quality moving: cache, retrieval, smaller model, measured on the evaluation set. Weekly supervision installed, then operated by us or handed to your team with the repository and documentation.
Interchangeable models behind an orchestration you own.
Models change name and price every six months; your instructions, limits, evaluation set and log stay. We choose the model per function, language, cost and hosting location.