Skip to content
Request a free call
About

Pooya Golchian

Pooya Golchian is an AI Transformation Lead in Dubai who ships agentic AI into production.

Read the writing, or get in touch.

Since 2009

The work, in order.

My first paid code shipped in 2009, freelance work that put me directly into client deadlines while I was still studying. I went on to earn a Master of Science in Software Engineering, which sharpened the discipline behind every system I have built since. The first full-time product work was frontend architecture across Angular, then Vue, then React. I built admin panels and customer-facing apps for an e-commerce company, a content company, and the InsurTech market in Iran. Different stacks, same job. Turn a designer's intent into something that loads fast on the worst connection in the room.

The middle stretch hardened the discipline. Five years across two regulated FinTech platforms, rising from senior engineer to leading a fifteen-person engineering pod. Real-time financial dashboards, a shared component library in React and TypeScript, and the SSR and accessibility work regulated products require.

A senior stretch at a Dubai payments platform was where the delivery muscle locked in. Payment gateway integrations, fraud and risk systems, and a shared engineering platform across a growing product line. I ran accessibility and localization to WCAG 2.1 because regulated markets don't ship without it.

Three years ago the AI work began in earnest. Not as a side topic, as the centre of gravity. I architect production agent systems with LangGraph and Vercel AI SDK. I design Model Context Protocol servers that expose internal data to AI clients safely, and build RAG pipelines that survive document churn. I write eval suites before shipping. I worry about token economics the way I used to worry about CDN costs. I self-host models on customer infrastructure for industries where data residency is the only acceptable answer.

The work stays full-stack. AI engineering pairs with the team's existing stack. That means Next.js App Router, Node.js services, monorepos and durable workflows. It also means the integration layer that makes a feature feel like one product, not three bolted together. I write code my colleagues can own after I leave.

From 2024 to 2026 I worked as an independent AI consultant. Owning delivery end to end is why I lead AI transformation instead of only writing code. The job is to move a team from AI strategy to a shipped product it owns, not just the code path.

Industries from fintech to insurtech to luxury retail to live gambling to AI-native B2B SaaS. The job at every one was senior engineering, shipped to production under deadline.

Timeline

Where the time went.

Newest role first.

  1. AI Transformation Lead & Forward Deployed Engineer

    FinTech / BNPL / LendTech company

    • Built the BNPL business end to end from scratch, then rose to lead its AI transformation. Wrote the backend in Java and Spring Boot across 18 microservices on PostgreSQL. Added Python and FastAPI services for the scoring and inference workloads and Go for the high-throughput paths. Built the customer and admin frontends in Next.js and React.
    • Made the call to leave the microservice estate behind and led the migration myself. Applied the Strangler Fig pattern to move traffic service by service into a Turborepo monorepo with a consolidated backend and a Backend-for-Frontend layer. New infrastructure, 35% less duplicated code, and a codebase shaped so agents could work in it.
    • Designed the agentic development lifecycle that estate now runs on, built on AWS's AI-DLC. Machine-readable specs drive Claude Code agents through generation, review and test behind evaluation harnesses and approval gates. Founded the practice, trained 25 engineers on subagent design and MCP tooling, and cut cross-team delivery time by roughly 40%.
    • Built the fraud, KYC and risk stack running across ten products. Trained and deployed supervised models in Python over engineered feature pipelines for risk scoring and transaction monitoring. Then combined them with third-party identity and fraud signals behind an agent fleet wired through MCP servers. Every automated decision clears a human review queue and writes an explainable trace, and fraudulent account creation fell 30%.
    • Carried the operating model past engineering, which meant sitting with each function rather than shipping a tool. Interviewed and onboarded 15 people across HR, Finance, Product Design, Marketing and Sales onto Claude Code with skills built for their own daily work. A shared knowledge base backs the skills so context travels between teams.
    • Built HazelPay on AWS with SST v3 and Next.js, unifying admin, dashboard and customer portals onto one platform. Built Pinapl (pinapl.ai), an AI marketing agent for Shopify merchants that runs abandoned-cart and reactivation campaigns autonomously under human approval.
    JavaSpring BootPostgreSQLNext.jsReactTurborepoClaude CodeMCPAWS

  2. Forward Deployed Engineer

    Independent AI Consulting

    • Embedded with UAE enterprise clients as their lead engineer, working on their infrastructure and inside their compliance constraints. Stood up NVIDIA GPU inference on customer premises. Fine-tuned Arabic models, Jais and Falcon, for organisations whose data cannot legally leave the country under PDPL.
    • Built on-premise Retrieval-Augmented Generation pipelines in Python answering in under three seconds. Responses are citation-first, so a client reviewer can check any claim against its source. Handed each deployment over to the customer's own team.
    • Built the batch and streaming data pipelines these systems run on, moving client data through PySpark, pandas and Polars into BigQuery and PostgreSQL. Tuned partitioning and clustering so analytical queries stay affordable at enterprise volume, and ran the surrounding services on GCP alongside AWS.
    • Designed and delivered Marteh, an agentic sales platform for GCC real estate, running voice agents in Arabic, English, Hindi and Urdu against live property data. Deal-cycle time fell 40% for the client.
    • Delivered a bilingual invoicing platform with voice-to-invoice, OCR and UAE VAT compliance, on AWS ECS Fargate provisioned through Terraform.
    LangChainLangGraphRAGTerraformECS FargatePrismaPython

  3. Senior Software Engineer & Engineering Pod Lead

    InsurTech / LendTech company

    • Rose from senior engineer to leading a 15-person pod across two regulated FinTech platforms, owning technical direction, delivery standards and mentoring.
    • Shipped a component library of 120+ React and TypeScript pieces in Storybook, which lifted feature-delivery speed across the organisation by 40%.
    • Rebuilt the Next.js and Nuxt.js server-rendering pipelines on high-traffic insurance and price-comparison sites, taking Lighthouse performance from 55 to above 95.
    • Architected a real-time financial dashboard in Vue and Electron, streaming market data over WebSockets at sub-second latency under heavy concurrent load.
    Node.jsReactVueNuxt.jsStorybookElectronWebSocket

  4. Full-Stack Developer (C# & PHP)

    E-commerce company

    • Built the company's first React storefront and admin order-management panel, setting a frontend architecture that carried the business for five years.
    • Delivered catalogue, checkout and order-management flows across a C# (.NET) and PHP (Symfony) backend for a high-traffic store.
    ReactC# (.NET)PHP (Symfony)Angular 1jQuery
In demand in 2026

What I work with.

The field has moved fast since late 2024. I use these skills daily.

Agentic systems

  • LangGraph & multi-agent orchestration
  • OpenAI tool calling + structured outputs
  • Vercel AI SDK with streaming UX
  • Anthropic Claude with extended thinking

MCP & integration

  • Custom MCP server design
  • Auth, rate limits, and capability scoping
  • Tool surface testing & versioning
  • Claude Desktop and IDE clients

Evals & observability

  • Golden datasets + LLM-as-judge calibration
  • Braintrust, LangSmith, Helicone tracing
  • A/B harness for prompts and models
  • Drift detection on production traces

RAG & retrieval

  • Hybrid search with reranking
  • Citation-first response design
  • Chunking strategies that survive churn
  • pgvector, Weaviate, Pinecone

Inference & cost

  • vLLM, Ollama, on-prem serving
  • Token economics & model routing
  • Quantization, LoRA, QLoRA fine-tuning
  • Open-weight models (Llama, Qwen, DeepSeek)

Voice & multimodal

  • Realtime API & low-latency voice agents
  • Vision pipelines for document AI
  • Whisper, Eleven Labs, Deepgram
  • Image generation (Flux, Midjourney v7)

Full-stack delivery

  • Next.js App Router + Server Components
  • Node.js services & durable workflows
  • Nx + pnpm monorepos
  • AWS via SST and OpenNext

AI safety & guardrails

  • Prompt-injection defenses
  • PII redaction & per-role access
  • Content moderation pipelines
  • Audit logs for regulated sectors
Practice

How I work.

Engineer first. I work inside the team, not over the wall. The deliverable is shipped code the engineers can own, plus the documentation and evals that prove it works.

No no-code tools. No vibe coding. No vendor lock-in baked into the architecture because it was easier in week one. Production AI is engineering work, and it will stay that way.

FAQ

Questions people ask about Pooya.

Common questions about Pooya's work, AI services, and how to start a project together.

  • Pooya Golchian is an AI Transformation Lead and AI Product Engineer based in Dubai, UAE. He has been shipping software since 2009, with the last three years focused on production AI. Pooya holds a Master of Science in Software Engineering.

  • Pooya Golchian treats AI transformation as hands-on engineering, not slideware. It means picking the workflows worth automating, then building the agent systems and RAG pipelines that run them. He self-hosts models where data residency matters, and the team that runs the system owns it.

  • Pooya Golchian builds production AI systems. That covers LangGraph and Vercel AI SDK agents, Model Context Protocol (MCP) servers, and RAG pipelines that cite sources first. It also covers evals using golden datasets and LLM-as-judge, self-hosted models, and full-stack Next.js on AWS via SST and OpenNext.

  • Pooya Golchian lives and works in Dubai, UAE. His production work spans fintech, insurtech, luxury retail, live gambling, and AI-native B2B SaaS.