Otun M.

I build AI systems that survive contact with real users, real traffic, and real invoices — not demos that break in week two. Most automation work falls into two camps. There are no-code builders who can wire an n8n workflow but can't tell you what happens when the LLM returns malformed JSON at 2am. And there are cloud engineers who can deploy anything but have never sat with a business owner who just wants missed calls to stop costing them money. I work across both. WHAT I ACTUALLY BUILD ▸ n8n automation & AI agents — self-hosted (Docker, VPS, nginx, Let's Encrypt) or cloud. Multi-step workflows with proper error branches, retries and dead-letter handling. Telegram, Slack, WhatsApp, Gmail, Google Sheets, CRM and webhook integrations. ▸ AI voice agents — inbound and outbound. Retell AI + Twilio (UK numbers) + GoHighLevel. Missed-call-to-text recovery, appointment booking, qualification and routing. Full stack: number provisioning, prompt design, call transfer logic, CRM writeback. ▸ RAG systems that don't hallucinate — pgvector on Postgres, hybrid dense + BM25 retrieval, LangChain orchestration. My preferred pattern is two-stage: deterministic SQL or structured retrieval first, then grounded synthesis, so every answer traces back to a source row rather than to the model's imagination. I've published a peer-reviewed paper on exactly this architecture. ▸ Production LLM applications — FastAPI + Next.js, streamed responses, Clerk auth, Stripe subscription gating. Deployed to Vercel for speed or AWS for scale: Lambda, App Runner, ECR, S3, API Gateway, CloudFront, Route 53. Amazon Bedrock (Nova, Claude) where the client needs data to stay inside their AWS account. ▸ Infrastructure that isn't clicked together by hand — Terraform with isolated dev/test/prod workspaces, GitHub Actions CI/CD using OIDC (no long-lived AWS keys anywhere), CloudWatch dashboards for token spend and latency, budget alarms so nobody gets a surprise bill. ▸ MCP servers and multi-agent orchestration — tool-calling agents with human-in-the-loop checkpoints at the steps where a wrong answer is expensive. WHAT MAKES THE DIFFERENCE Anyone can get an LLM to produce something plausible. The engineering is in making it produce the *same* thing tomorrow, proving where the answer came from, and knowing what it costs per call. So I build with that from day one: structured outputs with schema validation, retrieval that carries citations, eval sets before prompt tuning, retry and fallback paths on every external call, and cost observability wired in rather than bolted on afterwards. I've written and published research on production AI architecture and audit-trail design, and I've shipped multi-tenant SaaS on this stack — Next.js BFF, FastAPI compute API, queued CPU/GPU workers on AWS Batch, usage-based billing. That's also why I'm comfortable on projects where the output gets checked by someone other than the person who ordered it: compliance reviews, scientific workflows, financial reporting, anything with an audit trail. HOW I WORK I scope before I build. You get an architecture note and a fixed number before I touch a node. I write down what the system will refuse to do, not just what it will do — that's usually the part that saves the project. Handover includes documentation and a walkthrough recording, because an automation only I can debug is a liability, not an asset. Based in the UK. Comfortable working async across US and EU hours. If you're not sure whether your problem is an n8n workflow, a proper agent, or just a better-written Zap — message me and I'll tell you straight, even if the answer is "you don't need to hire anyone for this."