DXWand, 2019 to now. Senior engineer, then formal technical lead for a five-person cross-functional team from 2021 to 2023, then hands-on architect across the AI platforms.

Context

Enterprise customers, including Gulf-region government ministries, wanted answers grounded in their own documents through the channels their users already lived in. The company needed one retrieval platform it could ship to many customers rather than a bespoke build per contract.

The constraint

Every customer brought a different corpus, a different language mix, and a different tolerance for a wrong answer. The platform had to be tunable per customer without becoming a consulting project per customer.

The decisions that mattered

  • Auto-RAG. I architected and solo-built the MVP and shipped it to production for 7 to 10 enterprise customers. Later I contributed to the production Python rebuild: LangChain and LlamaIndex orchestration, Weaviate and Azure AI Search retrieval, and automated experimentation across retrieval strategies and prompt templates, so tuning became a measured run rather than an argument.
  • MerlinX. The flagship agentic conversational SaaS built on top of Auto-RAG: tool calling and multi-step reasoning, a .NET backend, a Svelte and TypeScript frontend, a Python AI layer, and Meta channels (WhatsApp, Instagram, Facebook). In production with about ten active customers.
  • Model economics. I found the orchestrator model was 75% of workflow cost, then proved the cheaper option was not worse with a frozen-suite comparison at 95% confidence. The rollout cut cost per conversation by 73% with no measurable quality loss across 840 agents, with a guarded apply and a one-command rollback.
Channels feed MerlinX agents, which use Auto-RAG retrieval and tools; a frozen evaluation suite gates every model and prompt change.Channelsweb · WhatsApp · Instagram · FacebookMerlinX agentstool calling, multi-step reasoningAuto-RAG retrievalWeaviate · Azure AI Search · customer documentsTools and integrationscustomer systems, actions, hand-offsFrozen evaluation suite73% lower cost per conversation, 840 agents
Retrieval underneath, agents on top, and a measured gate between any change and production.

Shipping it safely

  • Release engineering. Releases ship in gated waves. UAT runs the identical image artifacts, verified by digest rather than rebuilt look-alikes, and nothing promotes until post-deploy assertions pass and a soak holds against the baseline error rate. Waves ship rollback-free, including unattended overnight runs.
  • Governance at estate scale. I brought 35 Azure DevOps projects and about 224 repositories under one standard: an estate-wide branch-policy audit, then mandatory review gates on tier-one repositories, protected tags, a hardened deploy-render gate, break-glass access, and a scheduled drift audit. Every blocking finding in an independent security review was cleared.

Result

One retrieval platform serving 7 to 10 enterprise customers, an agentic SaaS with about ten active customers built on it, and a 73% cut in cost per conversation that was proven before it was rolled out. Figures are from the August 2026 CV.

Stack

Azure (App Services, AKS, Functions, Service Bus, Key Vault), C# and .NET, Python, LangChain, LlamaIndex, Weaviate, Azure AI Search, Svelte and TypeScript, Azure DevOps.

What it demonstrates

Architecture that survives many customers, the discipline to prove a cost decision statistically before rolling it out, and release engineering that lets a small team ship without fear.

September 2, 2026

All case studies

More case studies