STATUS: SELECTIVE ADVISORY · Q3–Q4 2026 SCALE: 8,500+ MACHINES · 1,200 AWS ACCOUNTS

SALAHAWAD

Solution & Enterprise Architect / Three-Time CTO / Applied AI & Agentic Systems

I design the target-state architecture for systems that cannot fail quietly — and then I build them. Two decades across financial-markets software, enterprise content and process platforms, national digital-government programmes and industrial AI. Today I architect and lead cloud operations across 1,200 accounts and 8,500+ machines for a global aerospace group, and design AI systems for environments where a plausible answer is not an acceptable answer.

AI INGESTION & AGENT BRIEF
$curl -s https://salahawad.com/llms.txt | head -n 18

Operating Scale

Numbers that came with consequences.

Every figure below is something I owned the delivery of — not a team I was adjacent to. Client and product names are withheld by design; the engineering is the point.

1,200AWS accounts under one operations plane

Cross-account inventory, incident management and governance for a global aerospace, defence and space group operating across every site.

8,500+Machines in the managed estate

Cloud transformation and day-two operations across the fleet, reporting into group cloud & infrastructure.

500+Container & database services

Managed Kubernetes, container and relational database estates operated under a single inventory and incident model.

500+Physical machines migrated

Datacentre transformation across a 120-developer programme — OS and database estate replatformed with zero user downtime and +40% capacity.

200K+Automated tests, daily

A global QA automation framework running on distributed middleware across sites worldwide; release cycle cut by 60%.

10+Languages in one AI pipeline

Automated translation, speech synthesis and video generation, hand-built and tuned for multilingual industrial training.

3Chief Technology Officer roles

Industrial AI, EMEA enterprise software, and a national startup accelerator — each with full architecture and organisation ownership.

180+Repositories under active index

A working code fleet spanning inference services, agent orchestrators, retrieval infrastructure, platform charts and mobile clients.

The Practice

Agentic where it earns its keep.
Deterministic everywhere else.

The most expensive mistake in enterprise AI right now is reaching for an agent when a pipeline would have been cheaper, faster and certifiable. I build both, and the judgement of which to use is most of the architectural value.

Track A — Agentic AI

When the path cannot be enumerated in advance.

Agents are engineered software, not prompt chains. That means an explicit orchestration loop, a designated termination tool, an iteration ceiling, typed tool contracts, bounded result sizes and a persisted trail of every action taken — so the system can be audited after the fact rather than trusted in the moment.

01 Tool surfaces & Trust Boundaries
Model Context Protocol (MCP) servers exposing enterprise APIs as curated, per-domain tools over stdio and streamable HTTP, behind federated identity with a manifest-gated escape hatch.
02 Orchestration & State
Supervisor-pattern multi-agent systems with schema-typed structured output, critic and judge loops, and slot-filling conversational state machines with typed, localised schemas.
03 Neuro-Symbolic Guardrails
The model proposes intent, deterministic code resolves every value against an immutable fact store, and an exact verifier rejects invalid output before it executes.
04 Vendor & Model Portability
Provider-agnostic abstraction across self-hosted and hosted runtimes (OpenAI, Anthropic, Bedrock, vLLM), so switching model providers is a configuration change rather than a rewrite.
Track B — Classical & Applied AI

When the task is bounded and accuracy is measurable.

Most of the value being shipped in production today is not agentic at all. It is speech, language, retrieval and scoring — tasks with ground truth, evaluation harnesses and a number you can put in front of an auditor. This is the unglamorous half that actually survives procurement.

01 Speech & Conversation Intelligence
Real-time transcription with speaker diarization, voice-activity segmentation, domain-dictionary correction, hallucination suppression, and sentiment, risk, empathy and quality scoring — including dialect-level work in Levantine Arabic.
02 Hybrid Retrieval That Is Measured
Lexical BM25 fused with dense vectors by Reciprocal Rank Fusion (RRF), tunable chunking, cross-encoder reranking, and dedicated evaluation harnesses so retrieval quality is a verified metric.
03 Multi-Modal Language Pipelines
Neural machine translation, text-to-speech, and programmatic video generation composed into staged pipelines with human-in-the-loop gates designed in as a stage, not bolted on.
04 Sovereign & Air-Gapped Deployment
Quantised open-weight models on local inference runtimes, GPU-semaphore concurrency control, CPU degradation paths, and air-gapped packaging for estates with zero egress.

Interactive Architecture Tool

AI Architecture Suitability Matrix

Toggle your enterprise constraints below to calculate whether your workload requires a Bounded Pipeline, a Supervisor Multi-Agent System, or a Neuro-Symbolic Guardrail.

1. Execution Path & Step Predictability
2. Tool Blast Radius & Side Effects
3. Ground Truth & Evaluation
4. Latency Budget & SLA
RECOMMENDED ARCHITECTURE

Bounded Deterministic Pipeline

Your constraints indicate a deterministic pipeline with measurable ground truth. Do not use an autonomous agent loop here: a structured pipeline with explicit evaluation harnesses will be 10x faster, cheaper, and fully auditable by compliance.

Mandatory Architectural Control:

Implement Reciprocal Rank Fusion (RRF) with BM25 + dense vectors and automated regression benchmarks in CI.

System Topologies

Interactive Architecture Studio.

Explore production-tested architectural topologies for high-consequence environments. Click on any node to inspect trust boundaries, contracts, and failure modes prevented.

PRODUCTION TOPOLOGY

Agentic Multi-Agent & MCP Trust Boundary

Zero Trust · Defense-in-Depth
NODE TELEMETRY & SPEC

Model Context Protocol (MCP) Gateway

Curated tool surface exposing enterprise systems as typed, named per-domain JSON-RPC contracts over streamable transports, strictly decoupling raw execution rights from model access.

TRUST BOUNDARY ENFORCEMENT Tokens scoped to individual tools rather than user's blanket rights. Escape hatch gated by explicit manifest.
PRODUCTION FAILURE PREVENTED Prevents prompt-injection-driven SSRF and unauthorized write mutations across backend datastores.
tool_manifest.json SCHEMA / CONTRACT
// JSON-RPC Tool Contract
{
  "name": "query_inventory",
  "description": "Bounded query of cross-account asset metadata",
  "inputSchema": {
    "type": "object",
    "properties": {
      "account_id": { "type": "string", "pattern": "^[0-9]{12}$" },
      "resource_type": { "type": "string", "enum": ["ec2", "eks", "rds"] }
    },
    "required": ["account_id", "resource_type"]
  }
}

Capabilities

Eight systems I am hired to design.

Each of these is a discipline I have owned end-to-end — architecture on the whiteboard, code in the repository, and the organisation that has to run it afterwards.

01

Target-state & enterprise architecture

Current-state assessment, competing architectural options scored against explicit criteria — security, availability, ecosystem interoperability, usability, scalability, total cost of ownership — then a defensible decision and the roadmap to execute it.

  • Future-state assessment
  • ESB vs. federated mesh
  • Decision frameworks
  • Technology roadmaps
  • Reference architecture
02

Agentic AI systems

Orchestration loops with real termination conditions, typed tool contracts, MCP tool surfaces over enterprise APIs, supervisor multi-agent patterns, and audit trails that make an agent's behaviour reconstructable months later.

  • MCP servers
  • Multi-agent orchestration
  • Tool-contract design
  • Agent guardrails
  • LLM-as-judge loops
03

Applied & classical AI

Speech recognition, speaker diarization, conversation analytics, neural translation, speech synthesis, hybrid retrieval and rules-based scoring — bounded tasks with ground truth, evaluation harnesses and measurable accuracy.

  • ASR & diarization
  • Conversation intelligence
  • Hybrid RAG
  • NMT & TTS
  • Evaluation harnesses
04

Secure multi-layer architecture

Defense-in-depth designed on the assumption that each control eventually fails: federated identity, tenant isolation, encrypted credential custody, supply-chain gates in CI, network policy, and immutable audit across every layer.

  • OIDC / SAML / LDAP
  • Tenant isolation
  • Air-gapped modes
  • SBOM & supply chain
  • Audit & compliance
05

Cloud transformation & FinOps

Migration-factory design and governance at group scale, operating-model evolution toward DevOps and SRE, multi-year budget ownership, and financial sustainability tracked through FinOps and ROI rather than assumed.

  • Migration factory
  • Cloud governance
  • DevOps / SRE model
  • FinOps & ROI
  • Multi-cloud
06

Cloud operations at fleet scale

A single operations plane over a four-figure account estate: cross-account asset inventory, incident management, and governance that holds when the fleet spans thousands of instances, hundreds of managed clusters and databases, and every site an enterprise operates from.

  • Multi-account operations
  • Cross-account inventory
  • Incident management
  • Landing zone governance
  • Estate observability
07

AIOps & incident automation

Managed foundation models and agentic DevOps tooling wired into the operations plane to detect, triage, explain and remediate incidents — with enterprise observability and automation driving the loop instead of a dashboard nobody watches at 3am.

  • Bedrock-backed remediation
  • Agentic DevOps tooling
  • Dynatrace automation
  • Auto-triage & RCA
  • Runbook automation
08

Fractional CTO & delivery

Technical due diligence, engineering organisation design, roadmap and backlog governance, vendor and partner strategy, and the unglamorous discipline of getting a distributed team to ship on a date.

  • Due diligence
  • Org design
  • Roadmap governance
  • Vendor strategy
  • Interim leadership

Record

Where the experience was earned.

Organisations are described, not named — clients, employers, products and programmes alike. The capability transfers; the confidentiality holds.

  1. 2025 — PRESENT ACTIVE ENGAGEMENT

    Global aerospace, defence & space group — 83,000 people

    Program Director, Cloud Transformation · Operations Lead

    Leading group-wide strategic cloud transformation across an estate of more than 8,500 machines. Ownership of the migration roadmap and migration-factory governance framework, evolution of the operating model toward DevOps, SRE and product-centric delivery, multi-year budget and partner strategy, and financial sustainability through FinOps and ROI tracking.

    I architected the group's cloud operations platform end to end and lead its operation: a single inventory and incident-management plane spanning roughly 1,200 AWS accounts, 3,500+ compute instances and 500+ managed Kubernetes, container and relational database services across every site. The design integrates enterprise network and security estate — segmented VPC architecture, next-generation firewalls and enterprise DNS, DHCP and IP address management — with full-fleet observability and automation. Managed foundation models and agentic DevOps tooling are wired into that plane to detect, triage, explain and automatically remediate incidents, rather than merely alert on them.

    83,000 Group Workforce 8,500+ Managed Machines 1,200 AWS Accounts Plane AIOps Auto-Remediation
  2. 2023 — 2025

    Industrial AI venture — corporate-incubated

    Co-Founder · Chief Technology Officer & CIO

    Designed and built an AI-powered industrial training platform from the ground up: a secure, modular ecosystem deployed across Azure, AWS, GCP and on-premise air-gapped environments, multi-tenant SaaS alongside private deployments for sensitive organisations, aligned to ISO and GDPR requirements. Personally coded much of the AI pipeline — automated translation, speech synthesis and video generation across more than ten languages — while acting as CIO over IT governance, cybersecurity and client conformity assessment. Extended reality and digital-twin methods cut training time by half.

    10+ Language AI Pipeline Air-Gapped Sovereign AI Digital Twins & XR 50% Training Time Reduction
  3. 2020 — 2023

    EMEA enterprise software vendor

    Chief Technology Officer

    Owned the technology strategy and stack across regional offices of an EMEA digital-transformation software vendor, leading teams of project managers, functional consultants and engineers across three offices. Delivered enterprise content services, business process automation, deep archival and data-governance platforms for financial services, government and healthcare — including semantic search with recommendation, records and legal-hold classification, PII and PCI handling, streaming-backed viewers, multi-source API gateways spanning CRM, e-signature and mainframe systems, and a core microservice platform of gateway, registry, storage, notification and scheduling services.

    3 Regional Offices Semantic Search & Archival Core Microservices Platform
  4. 2019 — 2020

    National startup accelerator & incubator

    Chief Technology Officer · Program Director

    CTO across an accelerator and incubator ecosystem, and program director for internationally funded agri-food and cleantech programmes. Set the technical direction for startups building on mobile, web, AI and blockchain stacks, drove each cohort from proof of concept through MVP to market-ready product, and built partnerships with major cloud and enterprise vendors to underpin the ecosystem.

  5. 2019

    Independent consultancy — GCC

    Senior Digital Transformation Consultant

    Technical architecture and technology selection for e-government programmes, directing a 28-engineer delivery team and defining architecture plans built to reduce total cost of ownership, alongside marketing-automation programmes for regional banks and enterprises.

  6. 2009 — 2019

    Global capital-markets software publisher

    Senior Product Manager · Project Manager · Senior Developer

    A decade inside one of the largest capital-markets software publishers, ending as technical infrastructure product owner for the flagship trading, treasury, risk and post-trade platform. Owned the infrastructure roadmap across cloud, virtualisation, operating systems, compilers, databases and hardware; led the technical operation moving a monolith to SaaS on public cloud; ran a cross-functional transformation migrating 500+ physical machines and replatforming the OS and database estate with zero downtime and 40% more capacity; cut hardware cost 30% and certification time 60%; and delivered a data-masking product and a global QA automation framework executing over 200,000 tests daily.

    500+ Physical Machines Migrated 200K+ Tests Daily Zero Downtime Replatform
  7. 2007 — 2009

    Regional banking software firm

    Team Leader · Senior Software Developer

    Led delivery of an internet banking application integrated with core banking web services, architected as a three-tier system with message-broker load balancing, plus stock-exchange reporting integration and a card rewards platform.

  8. 2017 — PRESENT

    Developer community leadership

    Community Lead — Beirut

    Leading a developer community of more than 7,000 engineers, hosting 55+ code labs, hackathons and technical events, including the first TechCrunch event in the MENA region.

Working Stack

What I actually build with.

Compiled from the live code fleet rather than a wish list — the technologies appearing in systems I have designed, written or run in production.

AI & inference

Open-weight LLM serving · quantised local runtimes · OpenAI · Anthropic · Azure OpenAI · Gemini · Groq · provider abstraction layers · Model Context Protocol · multi-agent orchestration · structured output · LLM-as-judge · prompt systems · evaluation harnesses

Retrieval & data

Hybrid BM25 + dense retrieval · reciprocal rank fusion · Apache Solr · OpenSearch · Qdrant · pgvector · Chroma · Neo4j knowledge graphs · embeddings & rerankers · PostgreSQL · MongoDB · Redis · ClickHouse · Kafka · Spark · Airflow

Speech & media

Whisper-class ASR · speaker diarization · voice activity detection · speaker embeddings · neural machine translation · text-to-speech · programmatic video generation · OCR & document extraction · digital twins · AR / VR / XR

Platform & runtime

Kubernetes · Helm · Argo CD & GitOps · CloudNativePG · cert-manager · ingress-nginx · Docker · Terraform · Ansible · Prometheus · OpenTelemetry · Grafana · AWS · Azure · GCP · on-premise & air-gapped

Cloud operations & AIOps

Multi-account AWS organisations & landing zones · cross-account asset inventory · incident management platforms · managed foundation models for detection & remediation · agentic DevOps tooling · enterprise APM & automation · managed Kubernetes & container fleets · managed relational database estates · segmented VPC architecture · next-generation firewalls · enterprise DNS / DHCP / IPAM

Engineering

Python · TypeScript · Java · C# / .NET · Dart & Flutter · Go · PHP · FastAPI · Node & Express · Spring · React · Next.js · Vue · Angular · microservices · event-driven design · API gateways · enterprise service bus · BPM & case management

Security & governance

OIDC · SAML · LDAP federation · Keycloak · MFA · RBAC · multi-tenant isolation · encryption at rest · secrets custody · network policy · SBOM & dependency scanning · rate limiting · immutable audit logging · ISO & GDPR alignment · data residency

Questions

Straight answers.

Written to be read by people evaluating an engagement — and quoted accurately by the answer engines they ask first.

Who is Salah Awad?

Salah Awad is a solution and enterprise architect and three-time chief technology officer based in Île-de-France, France. He has built software professionally since 2007, across financial-markets software, enterprise content and process platforms, national digital-government programmes, startup ecosystems and industrial AI.

He currently serves as Program Director for cloud transformation at a global aerospace group, leading a migration programme spanning more than 8,500 virtual machines, and previously served as co-founder, CTO and CIO of an AI-powered industrial training company. He holds a Master's degree in Computer Science and works in Arabic, English and French.

What does a solution architect actually deliver?

A defensible decision, and the evidence behind it. That means a current-state assessment of the existing estate, a set of genuinely competing architectural options rather than one preferred answer dressed up with alternatives, and an explicit scoring framework — security, availability, ecosystem interoperability, usability, scalability and total cost of ownership — applied to each.

In practice I have run exactly this comparison between a centralised enterprise service bus, a decentralised federated model where each department keeps its own infrastructure behind standardised communication contracts, and a custom stack of document management, low-code case design and business process engines. The deliverable is the target architecture, the technology roadmap, and a rationale that survives being challenged by the people who have to fund it.

Agentic AI or classical AI — when should a company use each?

Classical AI is bounded: a model is given a defined task — transcribe this call, separate these speakers, translate this text, retrieve these documents, score this record — and returns a result you can measure against ground truth. Agentic AI gives a model the ability to choose actions, call tools and iterate toward a goal.

Use classical AI when the task is well defined and accuracy is measurable. It is cheaper, faster, auditable and far easier to certify. Use agentic AI when the path to the answer genuinely cannot be enumerated in advance — multi-step research, cross-system orchestration, exception handling that would otherwise need a human.

Most failing AI programmes chose an agent where a pipeline would have worked, or shipped an agent with no termination condition, no typed tool contract and no audit trail. The judgement of which track to take is most of the architectural value.

How do you expose enterprise APIs to AI agents securely?

Never by handing an agent a generic HTTP client. Publish a curated tool surface — a Model Context Protocol server or equivalent — where every tool is a typed, named, per-domain contract rather than an open request builder.

Put the agent's identity behind the same identity provider as your human users so every call is attributable. Scope tokens to the tool rather than to the caller's full rights. Gate any generic escape-hatch tool behind an explicit manifest. Bound and truncate tool results. Log every invocation with its arguments to an immutable audit trail. The trust boundary belongs between the model proposing an action and the deterministic code that authorises it.

What is secure multi-layer — defense-in-depth — architecture for AI?

It is designing on the assumption that every control eventually fails, and layering independent controls so no single failure is fatal.

For AI systems the layers are: network and deployment isolation, up to fully air-gapped operation where data cannot leave the premises; federated identity with multi-factor authentication and role-based access control; per-tenant data isolation; a constrained tool surface for any agent; deterministic verification of model output before it takes effect; encryption of data at rest and of stored credentials; rate limiting and abuse controls; supply-chain gates such as software bills of materials and dependency scanning in CI; and immutable audit logging across all of it.

Can large language models run fully on-premise or air-gapped?

Yes — and for regulated, defence, sovereign and industrial workloads it is often the only acceptable design. Quantised open-weight models on a local inference runtime handle a large share of real enterprise work: transcription, translation, extraction, summarisation, classification and retrieval-grounded answering, with no data leaving the perimeter.

The architectural effort goes into the provider abstraction, so the same application runs against a local runtime or a hosted API by configuration; a CPU-only degradation path for sites with no accelerator; concurrency control so a single GPU is not overrun; and packaging the entire stack for environments with no internet egress.

What does conversation and contact-centre AI involve in practice?

Considerably more than calling a transcription API. A production pipeline needs voice-activity detection to segment audio, speaker diarization to establish who spoke when, a speech recognition model tuned for the actual dialect in use rather than textbook language, domain-dictionary correction for the vocabulary specific to that business, a model-based correction and hallucination-suppression pass, and text normalisation before any analysis begins.

Only then does the interesting layer run: sentiment, risk, empathy and quality scoring, referral suggestions, and searchable conversation history. I have built this stack end to end for Levantine Arabic — real-time, speaker-identified, with a bilingual right-to-left operator interface and local model inference so recordings never leave the operator's infrastructure.

How do you operate a cloud estate spanning over a thousand accounts?

At four-figure account counts the binding constraint stops being any individual workload and becomes coherence: knowing what exists, who owns it, what changed, and what is currently broken.

That requires a single operations plane above the accounts rather than per-account tooling — a canonical cross-account inventory of compute, container, cluster and database resources; an incident model that survives ownership boundaries; and governance applied at the landing-zone level rather than negotiated account by account. The network and security estate has to be integrated into that same plane — segmented virtual networks, next-generation firewalls, and enterprise DNS, DHCP and IP address management — or the inventory is fiction. Only once that spine exists is automation worth adding.

What is AIOps, and can AI actually resolve incidents automatically?

Alerting is not operations. Once a canonical inventory and a single incident plane exist, managed foundation models and agentic DevOps tooling can be wired directly into them — to correlate signals across the estate, triage an incident to its owning service, explain the probable root cause in language an on-call engineer can act on, and execute a bounded remediation rather than paging a human to run a runbook by hand.

The engineering discipline is identical to any other agentic system: bounded action surfaces, deterministic verification before anything changes, and an immutable audit trail of every automated action taken. An agent that can restart a service must not be able to delete one.

What kind of engagements do you take on?

Target-state architecture and technology roadmaps. Agentic and classical AI system design, and review of AI architecture that is already in flight. Secure multi-layer platform architecture for regulated environments. Cloud transformation and migration-factory programme design, including FinOps and ROI governance. Fractional or interim CTO engagements covering technical due diligence, engineering organisation design and delivery governance.

Engagements run in English, French and Arabic, across Europe, the Middle East and North America.

Contact

Tell me what has to survive.

The most useful first message describes the constraint, not the technology: the audit you have to pass, the outage you cannot have, the data that cannot leave the building, or the date that will not move.

Tip: Press Ctrl+Enter to send

Direct to Salah Awad's inbox. No automated marketing sequences, no third parties.