Artificial Intelligence

A useful PoC runs in 3 to 4 weeks. Full production (integrated into your IS, with monitoring and guardrails) takes 2 to 4 months depending on complexity. We ship in increments, never in tunnels.

Not necessarily. We can host on your cloud, on a Moroccan sovereign cloud (OCP), European (OVH, Scaleway) or American. For small volumes, a single A100 80GB is enough. Very large models can also be called via API (Claude, GPT-4o) through a zero-retention proxy.

We use natively multilingual models (Qwen 2.5, Mistral Large) and fine-tune on Darija corpora when relevant. For speech, we combine Whisper with post-processing specific to Moroccan Arabic.

Yes. With a self-hosted private LLM, no data leaves your infrastructure. Logs are encrypted and audited. We sign NDAs and can host in fully air-gapped environments if required.

We define metrics at scoping time: precision, recall, hallucination rate, user satisfaction. We use RAGAS, LLM-as-judge, and manually-curated test sets. Everything is reported in a dashboard you can consult at any time.

Yes, it's one of our specialities. We use LoRA/QLoRA to keep costs down. For a 10,000-example dataset, fine-tuning takes 12-24h and costs about €300-800 in compute.