Matin Labkhandagh
Book a call

Service

AI automation development

AI automation development means building the tools, orchestration and guardrails that let an AI operator run real business workflows instead of only talking about them. I, Matin Labkhandagh, build it with MCP servers, LangGraph, Tool Calling and FastAPI or Node.js against your existing database and channels, as in AiMatin, where an AI operator runs coupons, popups, funnel automations, reports and SMS workflows.

Book a 20-minute technical callSend a written brief

When you need this

  • Operations run through a few people who know the admin panel by heart, and every campaign, report or coupon waits for one of them.
  • You tried a chatbot that answers questions but cannot act: it cannot create the coupon, send the SMS or pull the report it is describing.
  • Marketing automation is spread across scripts, cron jobs and a spreadsheet, and nobody can say what ran last night or why.
  • You want an AI to operate internal tools, but handing a model raw database access is not acceptable and you need tool contracts, consent and an audit trail instead.
  • Leads arrive from several channels and follow-up depends on someone remembering to send the message.

What you get

  • Workflow inventory and tool design: which workflows the AI operator may run, each with typed inputs, outputs, permissions and the conditions under which it must ask before acting.
  • MCP server, in Node.js or Python, exposing those workflows as tools against your existing API and database, so any MCP-capable client or agent can operate them.
  • LangGraph orchestration with state and retry handling for multi-step workflows, so a failed step resumes rather than restarts.
  • Tool-calling hardening: argument validation, idempotency, consent for sensitive actions, and a ledger of every tool call.
  • Channel integrations as needed: SMS, Telegram, Bale, email, or your CRM, behind the same tool layer.
  • RAG integration where the operator needs your documents, catalog or past reports as context for a decision.
  • Evaluation tests that replay real workflows end to end and serve as the acceptance criteria for the sprint.
  • Docker deployment and documentation, including how to add a new workflow without me.

Outcomes

  • An AI operator that executes workflows through typed tools with explicit permissions, rather than free-form access to your systems.
  • Every automated action recorded, so what ran, when, on whose instruction and with what inputs is answerable.
  • Marketing, reporting and CRM-related workflows that run on triggers and schedules without a person in the loop, with a person able to step in.
  • Integrations with the channels your customers actually use, such as SMS, Telegram and Bale, behind the same tool layer.
  • Evaluation tests and documentation, so the automation can be changed safely after handover.

How it works

  1. 1

    Technical call

    A 20-minute technical call. You walk me through the workflows that eat your team's time and the systems they touch. I ask what must never happen automatically, who needs to approve what, and which channels your customers are on.

  2. 2

    Written diagnosis

    I write down the workflow inventory as I understood it, ranked by how much time it costs you and how safe it is to automate, and what your API and database expose already. The workflows that are risky to automate are named as such, not quietly included.

  3. 3

    Fixed-scope proposal

    Scope, timeline and price in writing. Most automation work fits the AI Agent Engineering Sprint, starting at $1,200 for 7–10 working days: a defined set of workflows exposed as tools, orchestrated, tested and deployed.

  4. 4

    Sprint and handover

    Tools first, against your real API and a test database, then orchestration, then the channel integrations, then the evaluation tests that replay the workflows. Handover is the MCP server, the orchestration code, the tests, Docker deployment and documentation your team can extend.

Proof: what I built

  • AiMatin is a solo-operated education and commerce platform I built and run. It is FastAPI and PostgreSQL behind a Next.js front end, with a custom Node.js MCP server that lets an AI operator run admin workflows: coupons, popups, funnel automations and reports. Marketing automation, reporting tools, Telegram and Bale integrations and automated SMS workflows sit behind the same tool layer. The site draws about 4,000 monthly organic unique visitors, roughly 90% of them from Google, and the popup funnels convert about 10% of visitors to leads, with the follow-up automated.

    Read the AiMatin case study

  • In OmidGPT I built the tool-calling runtime that this kind of automation depends on: an agentic tool-calling runtime with resumable tool turns, a tool-call ledger, per-tool user consent, MCP connectors and an OAuth / MCP relay, on a platform with 209 API endpoints and 50 database models. Those are the same patterns I use to make an AI operator safe to hand real permissions to.

    Read the OmidGPT case study

Questions about ai automation development

Which systems can the AI operator act on?
Anything with an API or a database you control: your admin backend, CRM, email and SMS providers, Telegram and Bale bots, payment and coupon systems. Each becomes a typed tool on the MCP server with its own permissions. Systems without an API can usually be reached through their database, with read and write scoped separately.
Can you work with an existing AI codebase?
Yes. If you already have an agent, a chatbot or scripts, I add the tool layer and orchestration around them rather than replacing them. If the existing code cannot be made safe to give real permissions, I say so in the written diagnosis and propose what to keep.
How do you keep LLM costs predictable?
With per-token cost accounting, semantic caching, complexity-based model routing and per-user quotas, as built in OmidGPT. For automation this matters less per request and more per schedule: a workflow that runs every few minutes needs a cost ceiling, and I set one in the orchestration.
What is the engagement process?
A 20-minute technical call, a written diagnosis with the workflow inventory, a fixed-scope proposal with scope, timeline and price in writing, then the sprint and a handover of code, documentation and tests. Tests replay the workflows and serve as acceptance criteria wherever practical.
Do you work project-based?
Yes. Automation work is scoped as a fixed set of workflows with a fixed price and timeline, usually the AI Agent Engineering Sprint. Adding workflows later is a new fixed scope, or your own team does it using the documentation.
Do you work with agencies?
Yes. Agencies bring me clients whose operations need an AI operator, and I build it white-label under the agency's name or alongside the agency's team. The MCP server and documentation are written so the agency can maintain them.