Saad LachhabAI Engineer · CTO
DXB --:--:--
[ SL / 2026 ] · Portfolio, 2026 edition 25.2048°N 55.2708°E · Dubai, UAE
Now: Senior Forward-Deployed AI Engineer @ DataRobot

I turnnoise intosignal.

AI engineer and 4× CTO. I design, build and ship agentic systems that governments and enterprises actually run in production, from the architecture whiteboard to the boardroom.

See the work Download CV
Scroll to decode
01The story so far

It started with noise. Hundreds of millions of sensor points from Airbus test flights, where one parasitic rattle could hold up a certification. I taught machines to hear it. The noise has changed since then. Legal codes, data rooms, telecom policies, government workbooks. The job is the same: find what matters in it, put it in production, and make sure someone pays for it.

Engineer

Software & AI engineering degree, UTC Compiègne. Aerospace R&D at Safran and Airbus.

Founder-CTO

Four times. Paris, London, Rabat, Casablanca. Products, teams, revenue.

Forward-deployed

Today in Dubai, embedded with government and telecom clients, shipping AI that survives audits.

02Impact, measured

What changed, in numbers.

Counted from production systems with real users and real contracts. Where a figure comes from a client report, I say so in the case study.

M.01Experience
0+
years engineering AI & softwareaerospace → enterprise
M.02Production
0
AI use cases shipped end-to-endUAE enterprise & government
M.03Customers
0+
customers served across my products and engagementsgovernment, telecom, finance, legal, private equity
M.04Velocity
0
codebases I actively shipped in the last 12 monthsproducts, client apps and platforms, in parallel
M.05Reach
0·6
countries and cities I've built fromParis · Toulouse · London · Rabat · Casablanca · Dubai
M.06Leadership
0
engineers led as CTO and Head of AI2 AI engineers mentored to promotion
M.07Revenue
$0K+
in enterprise deals as CTOincl. OCP Group, Big 4 & governments
M.08Recognition
0×
first prizes & national awardsMeta Llama MENA · AI for Tomorrow · Dyson
03Flight log

From the hangar to the Gulf.

2015 → 2026
WPT.00FL010
2015

Engineering school

UTC Compiègne · France

Software & AI engineering after ranking 1st in my prep class. National winner of the James Dyson Award in 2017.

#1prep-class rank
2017Dyson Award, FR
2018graduated
WPT.01FL050
2018

Software & Mechanical Engineer

Safran Aircraft Engines · Paris · Internship

Modal acoustic decomposition to study jet-engine noise. Research code that had to convince senior aerospace engineers.

R&Dengine acoustics
WPT.02FL120
2019

Acoustic Flight Test Engineer

Airbus · Toulouse

AI for parasitic-noise detection over hundreds of millions of in-flight sensor points. Certification of A350, A330 and A320.

80%verification cycle automated
3aircraft families
WPT.03FL200
2021

CTO

Lokalité · Paris

B2B SaaS + mobile app for food lifecycle analysis. Deep-learning carbon models over 15+ data sources, constraint-programming supply chains.

22enterprise accounts
8engineers led
WPT.04FL260
2023

CTO

DueDeal.ai · United Kingdom

An M&A due-diligence co-pilot for private equity: agentic RAG, financial benchmarking and a Neo4j graph of 2M entities.

12PE proofs-of-concept
96%retrieval accuracy
3engineers led
WPT.05FL310
2024

CTO

Juridia / Lawis · Rabat & Dubai

Four AI-native legal products: semantic search, assistant, contract lifecycle, regulatory watch. Sold to ministries and global firms.

70+institutions
92%research time saved
9engineers led
WPT.06FL350
2025

CTO

Ederest · Casablanca

Cloud-native ERP learning & digital-adoption platform on Azure with agentic AI and live practice sandboxes. Lead engineer and technical seller.

11K+users
$300K+pipeline generated
7engineers led
● CURRENTFL390
2026

Head of AI Engineering

Gradient Solutions · Dubai

Production AI for telecom, transport, banking and real estate across the GCC. Built the G-Brain co-pilot and the AI governance control plane.

34use cases in production
2flagship platforms
15engineers led
● CURRENTFL410
2026

Sr. Forward-Deployed AI Engineer

DataRobot · Dubai

Embedded with flagship UAE government and enterprise accounts, Dubai Police first among them. I own production AI applications from discovery to governed deployment.

E2Escoping → production
On-premair-gapped delivery
04Selected systems

Nine systems, up close.

Nine years of building taught me one rule: a system only counts once someone pays for it and relies on it. Here are nine that passed that test: the problem, what I built, how it’s wired and what it changed. Client work under NDA appears as a vendor-neutral version or as architecture.

01 / 09Gradient · 2026

G-Brain

Enterprise AI co-pilot · piloted for a UAE telecom

Problem

Enterprises want ChatGPT-grade productivity, but their documents, databases and policies can't leave the building.

What I built

A self-hosted AI workspace: chat with page-level citations, NL-to-SQL, an agent store with tools and skills, ontology-driven extraction, workflow automation, live Slides and Sheets editors and meeting notes, all multi-tenant and on-prem. For a UAE telecom operator, I architected the agent API: a LangGraph loop over FastAPI and Elasticsearch with role-based access.

12+product modules, one platform
~50REST paths, SSE streaming
323automated tests
TypeScriptFastifyDrizzlePostgres RLSRedisElasticsearchLangGraphYjsDataRobot LLM Gateway
gradientsolutions.ai ↗
G-Brain chat answer with page-level citations
COPILOT · CITED ANSWERS
ARCHITECTURE"Redis carries the job. Postgres carries the truth."
ON-PREM · MULTI-TENANT React appschat · docs Fastify APIauth · RBAC · SSE Agent runtimetools · skills · MCP Postgres + RLSsource of truth Redis → workersingest · embed LLM gatewayDataRobot ElasticsearchBM25 + vectors hybrid RAG
G-Brain workflow builder
WORKFLOW BUILDER
G-Brain agent store
AGENT STORE
02 / 09Gradient · 2026

AI Governance control plane

Sovereign AI for regulated industries · deployed on-prem

Problem

Banks and governments can't scale AI they can't audit. Every model, prompt and LLM call needs an owner, a risk score and a trail.

What I built

A control plane with a model registry and lineage, a guarded LLM gateway (PII, prompt injection, secrets, toxicity, tool allowlists), 5-factor risk scoring, staged approvals, incidents and audit-ready compliance mapping. Its motto: nothing ships unsigned and data never leaves.

5frameworks: EU AI Act, ISO 42001, SDAIA, DIFC, CBUAE
5guard types on every call
100%traced, metered, attributable
ExpressPrismaReact / ViteLLM GatewayOpenTelemetry-style tracesAzure OpenAI
Governance gateway trace graph of a multi-agent run
AGENT TRACE · TOKENS · COST · LINEAGE
GUARDED GATEWAYevery call · every guard
Agents Copilots Batch ML LLM GATEWAY PII redact Injection Secrets Toxicity Tool allowlist Metering Registry · lineage · risk score · approvalsAUDIT → EU AI Act · ISO 42001 · SDAIA ModelssovereignClaude · Qwen
LLM gateway overview: endpoints, calls, block rate
LLM GATEWAY · GUARDS
Compliance matrix across EU, Saudi and UAE frameworks
COMPLIANCE · EU · KSA · UAE
03 / 09Dubai Chamber · 2026

Hisaab حساب

Agentic accounting for UAE SMEs · AI Agents Championship, Dubai Chamber

Problem

UAE SMEs juggle VAT, corporate tax, e-invoicing and WPS payroll with spreadsheets, or with an accountant they see once a month. The books fall behind and penalties follow.

What I built

Eleven named AI agents run the books end to end: bank reconciliation, bookkeeping, PINT-AE e-invoicing, payables, VAT 201, corporate tax, payroll and WPS, audit and the month-end close. They work inside a deterministic double-entry kernel. Agents propose, the kernel disposes, and no model ever writes to the ledger. Each agent has its own autonomy level and AED ceiling, and anything uncertain escalates from rules to memory to the LLM to a human.

11AI agents, English & Arabic
205UAE legal parameters in the rules engine
0ledger entries written by a model
React 19FastifyDrizzlePostgres RLS + pgvectorBullMQStripeEN / AR
Hisaab home: cash, VAT countdown and 11 agents on duty
HOME · 11 AGENTS ON DUTY
ARCHITECTURE"Agents propose, the kernel disposes."
Bank feedsENBD · ADCB Bills & receiptsArabic · English Injection guardPII redaction 11 agentspropose only Ledger kerneluae-rules · 205 Human inboxapprove · correct BooksPostgres RLS escalation: rule → memory kNN → LLM → human
Hisaab Agent HQ: the finance team of AI agents
AGENT HQ · THE TEAM
Hisaab autonomy settings with AED ceilings per agent
AUTONOMY · AED CEILINGS
04 / 09Lawis · 2024–26

Lawis.ai

Legal intelligence for MENA & Africa · co-founder & CTO

Problem

Lawyers lose days searching laws, decrees and case law spread across official bulletins in three languages, and a wrong citation is unacceptable.

What I built

Four products: semantic search, an AI legal assistant, contract lifecycle management and regulatory watch. The core is a LangGraph state machine with a 9-route router, planner/replanner and an adversarial verifier that checks verbatim traceability and over-interpretation. It runs on hybrid retrieval over 136K documents from 22+ official sources.

70+organisations, 1,800+ users
92%research time saved
6jurisdictions, 3 languages
FastAPILangGraphAzure AI SearchHybrid BM25 + vectorDocling / OCRAKS · UAE NorthYjs CRDT
lawis.ai ↗
Lawis legal search with multilingual sources
SEMANTIC SEARCH · 14 LANGUAGES UI
AGENT GRAPHno citation, no answer
Rewritequery Router9 routes Plannersteps Executorhybrid search Replanner Synthesizecited draft Verifieradversarial check 136K docs · 9 indexes reject → revise
Lawis contract lifecycle board
CONTRACT LIFECYCLE
Lawis dashboard with sources by jurisdiction
SOURCES BY JURISDICTION
05 / 09Telecom · 2026

Document verification

Consumer post-paid onboarding · a UAE telecom operator

Problem

Every post-paid application is vetted by hand: Emirates ID, salary certificate, tenancy contract, utility bills. That's about 1,500 cases a day, with forgery risk on every file.

What I built

An 11-stage pipeline. Deterministic rules form the spine; constrained vision-LLM extraction and agentic adjudication fill the gaps where rules run out. It covers forgery forensics, cross-document consistency, reuse fingerprints and government-portal checks. Hard-fail rules override the agent, and that is enforced in code.

11pipeline stages, auditable
~1.5Kcases/day manual workload targeted
3outcomes: pass · human · reject
Node / TSQwen VLMZod schemasPDF forensicsDataRobot on-premReact
NDA Client system is confidential. Shown: vendor-neutral twin "VerifyAI".
DECISION PIPELINErules decide · models assist
VerifyAI verification brain: the 11-station workflow spine
11 STATIONS · DETERMINISTIC SPINE
06 / 09Keolis · Gradient · 2026

Transit & network agents

Dubai Metro demand & resource planning · executive AI training

Problem

Operations controllers have to re-plan trains and staff in minutes when events, weather or incidents shift demand. Executives needed to see what governed agentic AI really looks like.

What I built

A 7-node agent graph (intent, planner, investigator, constraints, allocator, critic, explainer) with six tools: forecast, events, weather, history, roster and optimizer. It self-critiques, explores counterfactuals and replans, and every call runs through the governance gateway. I demoed it live to Keolis leadership in an executive programme on ML, GenAI, agentic AI and governance. The same pattern powers Helix, a GCC telecom network-intelligence suite.

7agent nodes, critic loop
6tools incl. optimizer
3LLM modes: governed · live · offline
Node / ExpressSSEReactClaude via gatewayForecastingMapLibre
Transit IQ: Dubai Metro network forecast map
DUBAI METRO · LIVE CROWDING FORECAST
AGENT GRAPHplan → act → critique → explain
Intent Planner Investigator Constraints Allocator Critic Explain tools: forecast · events · weather · history · roster · optimizer revise
Helix autonomous network agents with pending approvals
HELIX · 6 AUTONOMOUS AGENTS · HUMAN APPROVALS
07 / 09Numeos · 2026

Data & AI foundation

ProPneu Group, France · tech lead, via Numeos

Problem

Every AI idea took around six months to reach production. The data lived in production MySQL databases nobody wanted to touch, personal data was mixed in, and EU AI Act obligations were coming.

What I built

A Google Cloud data warehouse in three medallion layers (raw, cured, business), fed by read-only micro-batch ingestion from production MySQL through Airflow. Personal data is pseudonymised at the raw layer, product master data is reconciled across sources, and EU AI Act model cards and GDPR controls are part of every layer. A governed FastAPI service behind API Gateway and a Vertex AI MLOps base open it to the business.

20working days, kickoff to foundation
0impact on production databases
17architecture decision records
GCPBigQueryAirflowCloud RunFastAPIAPI GatewayVertex AISQL
NDA Client data is confidential. Shown: the architecture.
numeostechnology.com ↗
MEDALLION ARCHITECTUREzero production impact
ERP MySQLread-only replica Shop MySQLread-only replica Catalogue feedssupplier data Airflow8 DAGs BIGQUERY RawPII pseudonymised on arrival · partitioned Curedclean · dedup · master-data matching Businessmetrics · product master data FastAPIAPI Gateway Vertex AIMLOps base Governance in every layer: least-privilege IAM · EU AI Act model cards · GDPR · 17 ADRs
08 / 09Ederest · 2025–26

Ederest

ERP learning & digital adoption · CTO

Problem

ERP roll-outs fail on adoption, not software. People need to practise on the real interface without breaking production.

What I built

An Azure-native LMS with hands-on ERP sandboxes, plus three Chrome extensions: Creator records walkthroughs, DAP plays guidance live, and Autopilot uses vision-LLM to fill forms with human-like timing. I was also the technical seller in front of enterprise executives.

OCP · Big4enterprise deals closed
11K+enterprise + B2C users
$300K+pipeline generated
Azure AKSFastAPIReact / Next.jsChrome MV3GPT-4.1 visionCI/CD
ederest.com ↗
Ederest Creator building a guided walkthrough on an ERP screen
WALKTHROUGH BUILDER · ON THE REAL ERP
Ederest: hands-on SAP practice simulation with AI-guided steps
ERP PRACTICE SANDBOX · SAP SIMULATION
09 / 09DueDeal.ai · 2023–24

M&A co-pilot

Due diligence for private equity · CTO, London

Problem

Due diligence means weeks of analysts reading data rooms and cross-checking companies, people and deals by hand.

What I built

A financial-intelligence co-pilot combining search, conversational AI and benchmarking. Under it sit a Neo4j knowledge graph of companies, people and deals, a 15M-vector store and agentic RAG with automated document pipelines.

2M · 8Mentities · relationships
96%retrieval accuracy
98%user satisfaction · 12 PE POCs
PythonFastAPINeo4jVector DBAgentic RAGReact
KNOWLEDGE GRAPH · LIVE SAMPLEhover a node
04.bMore systems

Eight more, in brief.

Smaller in scope, same standard: real users, a hard constraint, and a design choice I would defend.

Enterprise · 2026S.1

Document Redaction

True PII redaction for scanned PDFs, Word and PowerPoint. Rules catch Emirates IDs, IBANs and phone numbers, a small vision model finds names and addresses, and the glyphs are removed, not just covered.

Rules + VLM
hybrid detection
Nodemupdf WASMQwen3-VL
Numeos · 2026S.2

Numeos AO

Public-tender intelligence for Morocco. Portals are scraped every morning, each tender gets a 0–100 relevance score and a Go/No-Go analysis, and response files are drafted to the buyer’s template.

1–2.5 min
per drafted response file
FastifypgvectorGPT-5.5
Keolis-MHI · 2026S.3

Gateline

CV screening for Dubai Metro and Tram hiring. The model extracts evidence with verbatim quotes and deterministic code does the scoring, so changing the criteria never means re-reading a CV.

0
CVs re-read when criteria change
Evidence extractionDeterministic scoring
Telecom · 2026S.4

Retail Eye

In-store conversation analytics, built for 6–9K recordings a day. Quality gates, voice detection and Whisper large-v3, with denoising applied only when a file fails the gate, because blanket denoising raises error rates.

6–9K
recordings a day, by design
Silero VADWhisper large-v3DNSMOS
Skelter · 2025S.5

Audit automation

For French audit firms: accounting files from Sage, EBP, Cegid or anything else are standardised by agents, then trial balance, cash-flow statements and an AI analytical review are produced.

73%
less manual audit work
FastAPIMistral agentsAzure
Gradient · 2026S.6

Mizan

Brand reputation for a Dubai real-estate operator. Every Google review is analysed for sentiment and urgency, routed as a ticket, and answered by AI in the brand’s voice after a person approves.

100%
reviews analysed and routed
SentimentReply studioAlerts
CalJoy · 2026S.7

CalJoy

A photo-to-calories app for Arabic-speaking markets. The vision model only names the food and the grams; the calories come from a nutrition database, so no number is ever invented.

4.9 s
per photo, down from 11.7 s
ExpoNestJSLangGraph
UAE government · 2026S.8

Institutional assessment

An Arabic-first platform that drafts institutional assessment reports with an AI auditor beside each one. Every number in a proposed edit is checked in code against the sources. It runs fully air-gapped.

Air-gapped
on-prem, offline install
FastAPIQwen3bge-m3
04.cThe wall

The rest of the shelf.

Banking, telecom, real estate, retail, HR. Most of these started as a two-week build for a client meeting and ended up running somewhere.

CVMTelco churn · upsell
HelixNetwork intelligence
Auto-LoanCredit co-pilot
AstraProperty valuation AI
HisaabLive P&L
G-Brain SlidesAI deck editor
ComplaintsBanking AI agent
MizanBrand reputation
RegistryAI governance
NetPulsePredictive maintenance
KhalilTenant intelligence
LedgerRetail pricing AI
ATM NetworkExpansion scoring
OntologySchema-driven extraction
Hisaab VATForm 201 by Layla
Branch OpsStaffing forecast
LawisLegal assistant
CVICustomer value
Talk to dataGoverned NL→SQL
Chat-with-DocAR/EN citations
Guard profilesLLM gateway
ReconcileNoor matches the bank
HR AIEmployee assistant
LawisEverything's ready
AgentsAutonomous ops
G-BrainCited co-pilot
Transit IQMetro demand
05Method

How I build.

I start from the business value, plan for production constraints on day one, and treat the model as the least reliable part of the system. Click a layer.

05.bBeyond the code

Outside the editor.

// commercial

Closing deals

I've closed more than $830K in enterprise deals as CTO, including OCP Group (an $11B+ revenue enterprise) and PwC, won on technical credibility. I also brought 70+ institutional clients onto a legal AI.

$830K+
// leadership

Building teams

I've led 15 engineers as CTO and Head of AI, set architecture standards and delivery cadence, and mentored two AI engineers to promotion. I learned rigour in a 34-person Airbus certification team.

4× CTO
// executive

Talking to the board

I build AI transformation roadmaps with C-suite and directors, run executive AI training (ML, GenAI, agents, governance), and turn regulation into architecture under the EU AI Act, ISO 42001, SDAIA and CBUAE.

C-suite
Organisations I've shipped for2018 — 2026
Dubai Police GovUAE telecom operators TelecomDataRobot AI PlatformKeolis TransportOCP Group IndustryPwC ConsultingAirbus AerospaceSafran Aerospace Dubai Police GovUAE telecom operators TelecomDataRobot AI PlatformKeolis TransportOCP Group IndustryPwC ConsultingAirbus AerospaceSafran Aerospace
Grant Thornton AuditCGI IT servicesSGG Morocco GovMinistry of Transport GovCDG Prévoyance FinanceProPneu RetailGradient AI Grant Thornton AuditCGI IT servicesSGG Morocco GovMinistry of Transport GovCDG Prévoyance FinanceProPneu RetailGradient AI
05.cHonours
2025

Meta Llama · AI Design Drive

1st prize, MENA. Presented the architecture to regional industry leaders after a 3-month challenge.
2022

AI For Tomorrow

1st prize among 600+ AI innovation projects in France.
2018

European Innovation Academy

Fully sponsored fellow, Torino.
2018 →

Continuous learning

Certified by Stanford, Google, IBM, Microsoft and Johns Hopkins. Still studying every week.
2017

James Dyson Award

National winner, France.
Portrait of Saad Lachhab
ID · SL-0001
DXB · ONLINE
06Operator profile

Hello, I'm Saad.

I'm an engineer by training, aerospace by first job and founder-CTO by temperament. I've spent nine years at the line where research meets revenue: making models work on messy real-world data, then making them survive security reviews, procurement and real users.

I work in English, French and Arabic, which is why I build Arabic-first, RTL-native products for the Gulf by default. Right now I'm in Dubai, embedded with government and telecom teams, turning AI pilots into systems people rely on.

Engineering degree, Software & AIUTC Compiègne · 2015–2018
CPGE: Maths, Physics, CSRabat · ranked 1st in class
07Open channel

Let's talk.

I'm open to AI leadership, forward-deployed and architecture roles, and to ambitious problems in the Gulf and beyond.

saad.lachhab15@gmail.com COPIED ✓
LinkedIn ↗ GitHub ↗ Download CV