AI Architect in Dubai: Sovereign, Auditable AI for Regulated Banking

Take a look at the AI architect job ads that fill Google's first page for Dubai. LLM integration. AI and data platforms. "Scalable ML architecture." Those are the tools. Tools are NOT the architecture. The architecture is the set of decisions a regulator, an auditor and an operator have to live with after launch. For a UAE bank, four of them matter most. By far.

1. Where the data and the model run. The Central Bank of the UAE's guidance note on AI and ML (issued 11 Feb 2026) points licensed institutions to the rules on data being retained in the country (§5c). That decides hosting before anyone picks a model. First decision. Biggest decision. I covered the design in Architecting a Sovereign GenAI Platform for Regulated Banking in the UAE.
2. How models are routed and swapped. Frontier models change often. Very often. A provider-independent LLM gateway makes the model a configuration choice. And it's the natural home for version pinning, logging and the kill switch. See the gateway case study.
3. What the system may read and do. Retrieval sources, customer data and tool permissions decide the attack surface. Once an agent can act, a bad answer can become a bad ACTION. Very different problem. See auditable agentic workflows.
4. How you prove it's safe to launch, and to keep running. The same guidance note expects stress testing and validation, testing of automatic updates before implementation, periodic independent challenge, and "the clear and immediate ability, with human intervention, to cease use" of an AI system (§5d, §6b, §6c, §6f). It also expects customer disclosures in Arabic and English (§4b). An architecture that can't show this evidence is not finished. Not finished. WRONG!

Operating in the DIFC? Then Regulation 10 of the DIFC Data Protection Regulations adds duties for personal data processed by autonomous and semi-autonomous systems. I wrote about the architecture it implies in DIFC Regulation 10: The AI Architecture It Demands.

Four things. Every day.

• Sovereign AI platforms. Data residency, identity and audit designed in from the very first diagram. Bank data stays inside the country. Period.
• Model-agnostic LLM gateways. Routing across frontier and open-source models, with fallbacks and per-model cost and latency attribution. No application rewrites. None.
• Production RAG and agentic workflows. Retrieval over structured and unstructured banking data. Durable agent workflows with human-in-the-loop checkpoints and full observability. See Production RAG for a core banking suite.
• Defensive LLM evaluation. LLM red teaming for banking: finance-specific harm categories, severity scoring, go/no-go gates, and re-tests on every material change. Defensive ONLY. No exploits. No attack prompts.

Frankly, anybody can describe work. I'd rather SHOW it. Everything below is public and dated. Check it.

What	Where	Date
LLM Red Teaming for Banking: A UAE Go-Live Checklist (nine harm categories, severity scoring, five go/no-go gates, six re-test triggers, CBUAE mapping)	Hugging Face article	18 Oct 2026
LLM Red Teaming for Banking Severity Scorer, a free Gradio tool (Apache-2.0). Score findings, run the five gates, list missing evidence, export a markdown evidence pack. No prompts, no model calls, nothing stored	Hugging Face Space · Projects	18 Oct 2026
Architecting a Sovereign GenAI Platform for Regulated Banking in the UAE, a reference architecture	Hugging Face article	21 Jun 2026
DIFC Regulation 10: The AI Architecture It Demands	Hugging Face article	24 Jun 2026
Code and profiles	GitHub · Hugging Face · ORCID

All articles: Writing.

• Training. I've delivered many technical training sessions with BilgeAdam, for enterprises and for individual groups. Many. Live, hands-on, technical.
• Talk proposals under review. I submitted a technical talk on the open-source severity scorer and a CI gate to DevFest Istanbul 2026 (GDG Istanbul, 19 Dec 2026). I also submitted an English abstract on severity-weighted red teaming and sovereign deployment to the 5th IFC workshop on data science in central banking, organised by the BIS Irving Fisher Committee and co-hosted with the Central Bank of the Republic of Türkiye (Istanbul, 23 to 25 Mar 2027). Both are submissions. NOT confirmed talks. The organisers decide, and I'll update this page when they do.
• Topics I speak on (English or Turkish): LLM red teaming for banking, sovereign AI for banking in the UAE, model-agnostic LLM gateways, and turning red-team findings into a CI gate.

Organising an event? Get in touch.

Here's the deal. Three titles. Three very different jobs.

• An AI engineer builds and runs the pipelines: ingestion, embeddings, prompts, evaluations, deployment.
• An AI architect decides the shape those pipelines live in. Where data sits. Which model serves which task. What an agent may touch. How the system gets switched off. What evidence proves it's safe. In banking, the architecture decision is the one the board and the regulator WILL ask about.
• An AI consultant advises on what to build and in what order. For banks, the useful kind brings architecture and evaluation, not just a use-case list. A list alone? WEAK!

Folks, use these on any system. Including one a vendor built for you. Especially that one. They come from my go-live checklist.

1. Is the whole system scoped and inventoried: purpose, data, actions, model and version, hosting, languages, owner, kill switch?
2. Is there a finance-specific harm taxonomy mapped to obligations (AML and sanctions, conduct, data protection, OWASP Top 10 for LLM Applications)?
3. Is coverage multi-turn and multilingual, with Arabic and English at minimum?
4. Are findings scored by severity with an expert-written rubric, and does a human review every high-severity result?
5. Were the go/no-go gates agreed with risk and compliance before testing started?
6. Has the kill switch been tested end to end, with the time to disable recorded?
7. Are re-test triggers defined: model, prompt, retrieval corpus, tools, guardrails, languages?
8. Is there an evidence pack per release, with the test set kept access-controlled?

Seven out of eight is NOT done. Tick all eight. BTW, you can try the questions on your own findings with the free severity scorer.

18+ years of software architecture across banking, aviation, security and embedded systems. My roots are in reverse engineering and low-level systems work. Same precision, now on AI platforms. Education: Anadolu University. Languages: Turkish (native) and English. Based in Dubai, United Arab Emirates. More on the About page.

What does an AI architect do in a UAE bank?

An AI architect decides how LLM, RAG and agentic systems fit the bank's regulatory and operational limits. Where the data and the models run. How models are routed and swapped. What the system can read and do. How it is tested, monitored and switched off. In the UAE that design has to reflect the CBUAE guidance note on AI and ML of 11 February 2026. No shortcuts.

Who is Ercin Dedeoglu?

Ercin Dedeoglu is an AI architect based in Dubai. He designs sovereign, model-agnostic LLM, RAG and agentic systems for regulated banking. He works as an AI Architect at SBS, publishes on Hugging Face about LLM red teaming for banking and sovereign GenAI, and built the open-source LLM Red Teaming for Banking Severity Scorer. Views on this site are his own.

Does the Central Bank of the UAE require LLM red teaming?

Not by that name. The CBUAE guidance note issued on 11 February 2026 expects AI to be stress-tested and validated so it does not produce unsafe or unintended outputs. It expects automatic updates to be tested before implementation. It expects periodic challenge by independent third parties. A finance-specific red-teaming programme is a practical way to meet all three. This is my reading, not legal advice.

What is a sovereign AI platform for a bank?

A platform where the models, the retrieval data and the logs run inside the jurisdiction the bank is regulated in. Identity, audit and a kill switch are built in. And a gateway lets the bank change models without moving data abroad. Data stays home. Always.

Can I use the severity scorer for my own LLM system?

Yes. It is free and Apache-2.0 licensed. It does not call any model, holds no attack prompts and stores nothing you enter. It is an educational tool, not regulatory or legal advice.