FAQ
Questions a CTO asks first
Short, specific answers. If yours is not here, the 20-minute call is the fastest way to get it.
Working with an existing system
- Can you work with an existing AI codebase?
- Yes. Most engagements start inside a codebase I did not write. The audit is designed for exactly that: five working days to read the architecture, retrieval, tool calling, prompts, latency and token costs, and to hand back a prioritized roadmap. I work in your repository, on your branch conventions, with your CI.
- Can you work with LangGraph?
- Yes. LangGraph is my default orchestration layer for production agents. Jozveh-AI runs a LangGraph supervisor over eight agents with evaluation as code. If you already have a LangGraph graph, I harden state, retries, checkpoints and tool calling rather than rewriting it.
- Can you improve an existing RAG system?
- Yes. I start by measuring retrieval quality on a golden set, then fix the retrieval and grounding layers before touching prompts or models. In the Jozveh-AI bake-off, model choice was not the bottleneck; retrieval, reranking and grounding filters moved faithfulness more than swapping models did.
- Can you build MCP servers?
- Yes. I built MCP connectors and an OAuth / MCP relay for OmidGPT, with per-tool user consent and a tool-call ledger, and a custom Node.js MCP server that operates AiMatin's admin workflows. I build both the server (tools, auth, audit) and the client side (consent, resumable tool turns).
- Can you evaluate an AI agent we already built?
- Yes. The AI / LLM Technical Audit covers exactly that: architecture, retrieval, tool calling and prompts, plus hallucination, latency and token-cost analysis, an evaluation strategy and a roadmap. For a deeper harness (RAGAS, a claim-level LLM judge, golden-file tests) the follow-on is an engineering sprint.
Engagement and process
- How long does an AI engineering sprint take?
- The AI Agent Engineering Sprint is 7–10 working days, starting at $1,200. The Prototype → Production AI Sprint is 3–4 weeks, starting at $2,500. The AI / LLM Technical Audit is 5 working days, starting at $400. Scope, timeline and price are agreed in writing before work starts.
- What do you need to start?
- Read access to the repository, a way to run the system (or a recording of it failing), a sample of real inputs, and thirty minutes of the person who knows where the bodies are buried. If you have no evaluation data yet, that is normal; building a golden set is part of the first week.
- Do you work with agencies?
- Yes. I work as the AI engineering specialist alongside an agency's team or white-label under the agency's name, on the same fixed-scope terms. The agency keeps the client relationship; I keep the engineering accountable.
- Do you work project-based?
- Yes, exclusively. Every engagement is a fixed scope with a written proposal, tests as acceptance criteria wherever practical, and a handover your team can operate without me. I do not bill open-ended hours.
- What is the engagement process?
- A 20-minute technical call, a written diagnosis, a fixed-scope proposal, then implementation and handover with code, documentation and a runbook. You talk to the engineer who does the work throughout.
From the landing page
- What exactly do you do?
- I build and run production AI systems: LangGraph agents, RAG and GraphRAG pipelines, MCP infrastructure, and the evaluation and cost layer around them. I work as an independent contract engineer with founders, CTOs and AI product teams.
- Our agent works in the demo but fails with real users. Can you make it production-ready?
- Yes. That is the AI Agent Engineering Sprint: a fixed scope of 7–10 working days, starting at $1,200, covering LangGraph orchestration, state and retry handling, tool-calling hardening, MCP and RAG integration, evaluation tests, Docker deployment and documentation.
- Our RAG answers are wrong or hallucinated. Where do we start?
- Start with the AI / LLM Technical Audit: five working days, starting at $400. You get an architecture and retrieval review, a tool-calling and prompt review, hallucination, latency and token-cost analysis, an evaluation strategy and a prioritized implementation roadmap.
- How do you keep LLM costs predictable?
- With per-token cost accounting, semantic caching, complexity-based model routing and per-user quotas. I built and operate exactly this stack in OmidGPT, a multi-provider agentic AI platform.
- How do you measure whether an AI system is actually good?
- With evaluation as code: RAGAS, a claim-level LLM judge, post-generation grounding filters and controlled model comparisons, wired into tests that serve as acceptance criteria. Jozveh-AI, an eight-agent GraphRAG pipeline, runs 673 tests across 77 files.
- Do you work with companies outside Iran?
- Yes, remotely. I am based in Iran and work with clients whose compliance and banking arrangements permit contracting an Iran-resident engineer. I raise this on the first call so both sides can confirm feasibility before starting.
- What happens after the 20-minute call?
- You get a written diagnosis, then a fixed-scope proposal with scope, timeline and price. Implementation ends with a handover: code, documentation and a runbook your team can operate without me.