About
About Matin Labkhandagh
I am Matin Labkhandagh, a Production AI & Agentic Systems Engineer. I build and run production AI systems: LangGraph agents, RAG and GraphRAG pipelines, MCP infrastructure, and the evaluation and cost layer that keeps them reliable. Ten years in software, five of them in AI.
I work as an independent contract engineer with startup founders, CTOs, engineering leads and AI product teams, remotely, in English or Persian. The work is fixed-scope: a technical audit, an agent engineering sprint, or a prototype-to-production sprint.
What I focus on
- Agentic systems that survive real users: explicit state, checkpoints, retries, bounded tool calls and a ledger of what every agent did.
- Retrieval that can be measured: golden sets, reranking, grounding filters and evaluation as code, so a wrong answer is a failing test rather than a support ticket.
- Serving economics: per-token cost accounting, semantic caching, complexity-based model routing and quotas, built before the bill becomes the incident.
Experience
| 2025 – now | Founded the AiMatin academy (Persian prompt-engineering course) and built the AI-operated business platform behind it. Run OmidGPT and Jozveh-AI in production. |
|---|---|
| 2023 – 2025 | Designed and launched AI agent systems for Iranian and international businesses: RAG pipelines, support automation, automated content generation. |
| 2021 – 2023 | Specialized in large language models from the GPT-3 era; first real-project implementations. |
| 2015 – 2021 | Full-stack software development; began machine learning. |
Engineering philosophy
Evaluation before optimization. I do not tune a prompt or swap a model without a harness that says whether it helped. Jozveh-AI has 673 tests across 77 files; most of them encode what a correct answer looks like.
Boring reliability. Resumable tool turns, idempotent side effects, explicit retry budgets and a fallback provider. The interesting part of an agent should be what it does, not how it fails.
Cost is an architecture decision. Accounting per token and per user comes first; caching and routing follow from the numbers, not from a hunch.
Handover is the deliverable. Code, documentation and a runbook your team can operate without me. Tests are used as acceptance criteria wherever practical.
How I work
- 20-minute technical call. You walk me through the system and what is failing. I ask the questions that decide feasibility.
- Written diagnosis. What is actually wrong, what it would take to fix, and what I would leave alone.
- Fixed-scope proposal. Scope, timeline and price agreed in writing before any work starts.
- Implementation + handover. Code, documentation and a runbook your team can operate without me.
Tests are used as acceptance criteria wherever practical.
Location and compliance
I'm based in Iran and work with clients whose compliance and banking arrangements permit contracting an Iran-resident engineer. I raise this early so both sides can confirm feasibility before starting.
Systems I built and operate
Three production systems, designed, shipped and still running. The numbers are the ones I can stand behind.
OmidGPT — Multi-provider agentic AI platform
A multi-model AI assistant needed frontier models from multiple providers, real-time streaming, tool calling, voice, moderation, and sustainable serving economics.
- API endpoints
- 209
- Database models
- 50
- Lines of Python + TypeScript
- ~94K
- Commits
- 391
Jozveh-AI — Eight-agent GraphRAG pipeline
Turn heterogeneous source material into cited, structured study documents while maintaining faithfulness and predictable cost.
- Tests
- 673
- Test files
- 77
AiMatin — AI-operated business automation
A solo-operated education and commerce platform needed automation for marketing, reporting, CRM-related workflows, and operations.
- Monthly organic unique visitors
- ~4,000
- Organic traffic from Google
- ~90%
- Visitor-to-lead conversion, popup funnels
- ~10%