Skip to content
Brocode SolutionsAI Software Development

Production ML, UAE-resident pod

Notebook to production. Twelve weeks.

A named pod of senior ML engineers takes the model your team built and ships it - monitored, retrainable, and audit-ready against CBUAE and DFSA model-risk standards.

Book the 45-minute reviewMessage on WhatsApp
fraud_v3.ipynbexperiment

import sklearn

from xgboost import XGBClassifier

model = XGBClassifier(...)

model.fit(X_train, y_train)

# TODO: serve this somehow

# TODO: retrain? monitoring?

fraud_v3 - prodlive
Requests / min
1,247
p95 latency
42ms
Drift score
0.03
Last retrain
4d ago

Feature Store

Feast

Training

MLflow

Registry

MLflow Models

Serving

BentoML / Triton

Monitoring

Evidently + Prometheus

Notebook -> Production. 12 weeks.

  • AED 14M

    Fraud recovered in six months - UAE Tier-1 bank reference

  • 4.1 pp

    Gross-margin uplift - regional logistics-tech pricing model

  • 11 weeks

    Notebook-to-production median across last 18 engagements

  • 18/47

    Median readiness score - 22 UAE enterprises benchmarked

The gap

Nine prototypes. One production model. The audit committee wants to know why.

Three failure modes account for almost every stuck ML programme we audit in UAE enterprises. Each one maps directly to an article in the regulator's most recent circular.

No retraining pipeline

A model trained once on a Q1 snapshot is a model that silently rots. CBUAE Model Risk Management Standards expect a documented retraining cadence and trigger conditions; without that, the model is an audit finding waiting to be written up.

No monitoring

Notebook-grade scoring has no latency SLO, no drift detector, and no link to your Splunk or Datadog. DFSA SEMC guidance on algorithmic systems requires operational telemetry that an internal auditor can reproduce on demand.

No audit trail

Without a feature store, an experiment tracker, and a model registry, the lineage from a single decision back to the training data is reconstructed by hand if at all. That answer cannot be produced inside the regulator's response window.

The pod

A fixed-shape ML Delivery Pod, named engineers on the contract.

Six seats. UAE-resident or relocatable. One engagement contract. No partner-to-offshore handoff.

Pod composition

1 ML lead + 2 senior MLEs + 1 MLOps + 1 data engineer + 1 solution architect

Every name in the proposal appears in the standup, the commit log, and the regulator handover. We send full CVs with public GitHub, conference talks, and certifications before contract signature.

Engagement

Fixed-price

12 weeks, fixed scope. Optional run-phase SLA after handover.

Location

UAE-resident

On-site Dubai or Abu Dhabi, four days a week minimum during build phase.

Stack

Feast + MLflow + Airflow + BentoML + Evidently

Open-source spine. Portable. No hyperscaler lock-in.

Partners

NVIDIA Inception, AWS Advanced, Databricks, G42 Cloud, Snowflake

Same engineers, certifications across all four ecosystems.

The contract

One signing. Six names. Zero subcontractors.

The MSA assigns model artefacts, code, Feast features, and MLflow logs to your organisation on creation. The DPA is UAE PDPL and GDPR aligned, with EU SCCs pre-filled if any data ever crosses borders. Customer-managed keys are the default, not a premium tier.

The 12-Week Production Path

Four phases. Named deliverables. Regulator artefacts at every gate.

The default rhythm for any production ML engagement. Fixed-price option available; pilot-and-roll variants on request.

  1. Weeks 1-2

    Discovery and production-readiness baseline

    On-site discovery with your data team. We score your current pipeline against the 47-point readiness audit, agree the first use case, and lock the success metric with the sponsor and the model-risk owner.

    Deliverable: signed scoping document + readiness scorecard

  2. Weeks 3-6

    Prototype on your real data

    The pod stands up Feast, MLflow, and an Airflow DAG inside your VPC or G42 tenancy. We re-implement the candidate model with a documented training run, a held-out validation set drawn from your live traffic, and a challenger model for benchmark.

    Deliverable: tracked experiments + challenger comparison

  3. Weeks 7-10

    Production hardening with paired retraining

    BentoML or NVIDIA Triton serves the model behind a canary. Evidently AI and Prometheus stream drift, latency, and accuracy into your Splunk or Datadog. A paired retraining DAG runs on the same data contract so the second cycle is the team's, not ours.

    Deliverable: first production deployment + paired retrain

  4. Weeks 11-12

    Regulator-ready handover

    We produce the Model Risk Pack aligned to CBUAE Model Risk Management Standards and DFSA SEMC: data lineage, validation evidence, challenger results, fairness review, and the ongoing monitoring plan. Internal audit walks the artefacts before signature.

    Deliverable: Model Risk Pack + audit walkthrough

What buyers usually push back on

Three objections we expect on the first call, and how we answer them.

Objection 01

“We have been burned by consultancies that send a partner to kickoff and then we never see them again. Show me the actual engineers and their CVs.”

Every proposal is accompanied by a named-engineer CV pack with public GitHub, LinkedIn, conference talks, and certifications for the exact people who will be on your engagement. The same names appear on the contract. The same names show up in your standups. If a person rotates, that is a contract amendment, not a quiet body-shop swap.

Proof: anonymised UAE Tier-1 bank reference - fraud model shipped in 11 weeks, AED 14M recovered in the first six months, passed internal audit on first review with the same four engineers on the page from week one.

Objection 02

“Our data cannot leave the country. Can your team work entirely inside our VPC, on our SageMaker tenancy, or on G42 Cloud without ever extracting data?”

Yes. The pod is UAE-resident, contracted under UAE law, and the entire delivery rhythm assumes your data does not move. We commission MLflow, Feast, and the retraining DAGs inside your tenancy. No customer record, no aggregate, no embedding, ever lands on a Brocode-controlled bucket. The MSA and DPA we sign say so in writing.

Proof: anonymised GCC family-office reference - portfolio risk model running on G42 Cloud with full data residency, quarterly retraining, and customer-managed keys on every Feast feature.

Objection 03

“Our internal data science team will resist this. How do you knowledge-transfer instead of locking us into a multi-year retainer?”

The pod pairs with your team from day one. The MLOps engineer and data engineer sit beside your senior data scientists; the second retraining cycle is run by them, not us, while we observe. The Model Risk Pack template, the Feast feature definitions, and the Airflow DAGs all live in your repos. Our exit clause is one month, not twelve.

Proof: anonymised regional logistics-tech reference - the pricing pod handed over fully after 14 weeks; the in-house team has run the last 11 retrains without us, lifting gross margin 4.1 percentage points in five months.

Reference architecture

The Brocode Production-Ready Reference Architecture.

The same five-layer spine on every engagement. Each layer is portable across AWS UAE North, Azure UAE North, G42 Cloud, or on-premise NVIDIA estates.

Layer 1

Feature Store

Feast

Versioned offline + online features.

Layer 2

Training

MLflow

Tracked runs, signed lineage.

Layer 3

Registry

MLflow Models

Stage promotion + approvals.

Layer 4

Serving

BentoML / Triton

Canary + shadow scoring.

Layer 5

Monitoring

Evidently + Prometheus

Drift alerts into Splunk / Datadog.

Drift signals flow from Evidently and Prometheus into your existing Splunk or Datadog so on-call routes through the team that already knows the pager.

Engineers reviewing an MLflow run tracker and feature-store dashboard inside a UAE bank's secure workspace

Case studies

Three references the prospect can phone.

  • UAE Tier-1 bank

    Real-time fraud-scoring engine shipped in 11 weeks. AED 14M recovered in the first six months. Passed internal audit on first review. Pod size: 5.

  • Regional logistics-tech

    Dynamic-pricing model lifted gross margin 4.1 percentage points in five months. Internal team now owns the retraining cadence. Pod size: 4.

  • GCC family-office holding

    Portfolio risk model running on G42 Cloud with full data residency. Retraining quarterly. Customer-managed keys throughout. Pod size: 4.

How we compare

Big-4 partners, offshore body-shops, hyperscaler PSO, or hiring in - the honest read.

The four routes a CDO actually weighs when this budget is on the table. Where we are not the right fit, we say so.

CapabilityBrocodeBig-4 (Accenture, Deloitte, KPMG, EY)Offshore dev shopHyperscaler PSODIY hiring
Named senior engineers on the contractYes - CVs sent with proposalPartner + offshore subcontractorsRotating body-shopRotating PSO consultantsYou hire them
Fixed 12-week path to production6-month diagnostic typicalTime and materialsT&M, scope drift9-14 months to hire team
CBUAE / DFSA Model Risk Pack templateIncluded in every engagementOn request, extra fee
On-premise / G42 Cloud deliveryHyperscaler-firstHyperscaler-firstCloud-vendor lock-inDepends on hire
UAE-resident delivery pod100% of pod under UAE labour lawUAE partner + offshore buildKarachi / BangaloreUAE-region cloud, US-billed teamUAE hires only
Reference calls without an NDA gateThree customers reachable by phoneNDA-gated case studiesSanitised slide-wareVendor logos onlyN/A
Engineering-hour blended rate45-55% of Big-4 published GCC ratesBig-4 standard ratesLowest, no audit trailUS-billed, premium~AED 7-10M / year fully loaded
Knowledge transfer commitmentOne-month exit, repos with youMulti-year retainer pullCode in offshore reposCloud-managed services lock-inInternal-only IP

Free download

From Notebook to Production: the 47-Point Readiness Audit

A 28-page PDF and an Excel scoring template, built around CBUAE Model Risk Management Standards and DFSA SEMC guidance. The median UAE data science team scores 18 of 47.

  • Data foundation readiness
  • Modelling discipline
  • MLOps backbone
  • Risk and governance
  • Operational support
  • Scoring rubric and benchmarks - median 18/47 across 22 UAE enterprises

Instant download. No spam. Unsubscribe any time.

Questions we get on the first call

Frequently asked.

Everything below is what we already say in writing on the SOW. If the answer changes for your engagement, we redline it before signature.

Ask a different question

  • Every proposal is accompanied by a named-engineer CV pack with public GitHub, LinkedIn, conference talks, and certifications for the exact people who will be on your engagement. The same names appear on the contract. The same names show up in your standups. If a person rotates, that is a contract amendment, not a quiet body-shop swap. Proof: anonymised UAE Tier-1 bank reference - fraud model shipped in 11 weeks, AED 14M recovered in the first six months, passed internal audit on first review with the same four engineers on the page from week one.

Talk to the pod

A senior ML engineer replies within one business day.

Tell us the use case, the regulator, and the residency constraint. We come back with named engineers, a draft 12-week plan, and three references you can phone.

Direct WhatsApp: +971 50 761 2213

Email: hello@brocode.ae

HQ: Al Maryah Island, ADGM, Abu Dhabi

Quote request

Book a 45-minute production-readiness review with a senior ML engineer

A senior MLE on the Brocode pod reviews your current pipeline, names the blockers, and replies within one business day.

Prefer chat? Message us on WhatsApp — we'll see it within working hours.

Continue exploring

Related capabilities and stories

Book the reviewWhatsApp