Arabic NLP that works on Khaleeji.

Objection 01

“Show me the Khaleeji dialect accuracy - not MSA, not Egyptian. We have been burned by Arabic-supporting vendors who only handle Modern Standard.”

The Pre-Contract Arabic NLP Benchmark exists to answer exactly this. Send us 5,000 anonymised utterances from your real conversation logs; we run the Brocode stack against them and share the accuracy report - broken out by dialect (Emirati, Saudi, Kuwaiti, Levantine, Egyptian), by channel, and by code-switch density. The Khaleeji dialect head is trained on a proprietary 4.2M-utterance corpus collected by UAE-resident Emirati, Saudi, and Kuwaiti annotators.

Proof: anonymised UAE telco reference - Khaleeji intent classification lifted from 64% to 91%, contact-centre Arabic deflection increased from 18% to 46% within four months.

Objection 02

“Our customer data cannot leave the country. Your stack must run on-premise or on G42 Cloud, and the annotation pipeline must be UAE-resident.”

The entire Arabic Language Stack runs without sending a single customer utterance to a US-billed API. Default deployment targets are on-premise, G42 Cloud, or AWS / Azure UAE North - chosen with your security team. The Khaleeji Dialect Annotation Lab is staffed by UAE and Egypt-resident annotators under ISO 27001 controls, with full audit trail on every labelled sample. PII redaction happens on your perimeter before any model call.

Proof: anonymised UAE bank reference - Arabic complaint-routing accuracy at 94%, full PII-redaction pipeline, audit-trail signed off by the bank's Model Risk Officer, all on-premise.

Objection 03

“We already pay an LLM API provider. Why would we run our own model instead of just prompt-engineering GPT-4 with Arabic instructions?”

Three honest reasons. Dialect: frontier APIs ship MSA-flavoured Arabic with degradation on Khaleeji - we ship a Khaleeji dialect head trained on Gulf-resident annotators. Data residency: their inference runs outside the UAE, ours does not. Economics: their Arabic tokens are 2-3x English tokens by tokenizer, so per-call cost crosses over a self-hosted Brocode stack around 4-6M calls per month for most clients. We will show you the crossover analysis for your volume in writing.

Proof: anonymised federal entity reference - Khaleeji voice-note transcription Word Error Rate of 11.2% vs ~31% on hyperscaler defaults, on-premise, with per-call cost ~70% lower at the client's monthly volume.