Guide · AI development

AI software development: how AI products are actually built

The model is the smallest part of an AI product. What makes it useful and trustworthy is everything built around it: how it reaches your data, what it is allowed to do, how you know it works, and what happens when it does not. This guide explains that work, so you can judge it when you buy it.

Updated 3 September 20269 min readBy Noxira

01The anatomy of an AI product

Model access
A frontier model through its API (Claude and comparable models), or an open-weight model run privately where data residency or cost asks for it.
Retrieval
Your documents and records indexed so the model answers from your facts — usually PostgreSQL with pgvector, with citations returned.
Tools
The functions the model may call to read and write your systems, each with a permission boundary enforced in code.
Orchestration
The code that runs the loop: prompts, tool calls, retries, fallbacks. Plain, testable code beats a framework for most products.
Evaluation
Real examples with correct outcomes, run automatically on every change. The difference between engineering and hoping.
Guardrails
Limits, checks, approval steps and safe failure when the model is unsure.
Observability
Traces of every call, cost per task, quality over time, alerts when either drifts.
Interface
Where people meet it: chat, WhatsApp, voice, or a feature inside an existing app.

02The process, from pilot to production

  1. Frame one task with a measurable right answer. Reject tasks without one.
  2. Ground the model in your data and write the evaluation set — fifty real examples is a good start.
  3. Build the smallest system that passes the evaluations with a person approving.
  4. Guard it: limits, fallbacks, checkpoints.
  5. Deploy in your cloud with cost and quality visible.
  6. Operate: read the traces, tighten the loop, widen the scope.

03The stack

TypeScript and Python; PostgreSQL with pgvector for retrieval; LangGraph or hand-written orchestration; serverless GPU (for example Modal) when a model must run privately; deployment on AWS, Vercel or your infrastructure; evaluation suites kept in the repository beside the code. Chosen so your own team can maintain it.

04Costs and timelines

A pilot is about four weeks; a production system about eight. At Noxira they start from QAR 30,000 and QAR 70,000 respectively, fixed-price after scoping. Running costs are per model call and should be visible in a dashboard from the first day.

05Choosing an AI development company

  • Ask to see an evaluation suite from a past project. If there is none, the work was not measured.
  • Ask where the system runs and who owns the accounts. The answer should be: yours.
  • Ask what happens when the model is wrong. There should be a specific answer.
  • Ask whether they run AI in production themselves. Operating it teaches things building it does not.
  • Prefer a team that designs the interface and the data model as well as the prompts — an AI feature is a product.

Questions we get asked

Do we need our own model?

Almost never. Frontier models through an API, grounded in your data with retrieval, outperform a custom-trained model for nearly every business task, at a fraction of the cost.

What is RAG?

Retrieval-augmented generation: searching your own documents for the relevant passages and giving them to the model with the question, so it answers from your facts and can cite them.

How do you know an AI system works?

With an evaluation suite — real examples with correct outcomes — run automatically on every change, plus traces and quality metrics in production.

Can AI software be built in Qatar?

Yes. Noxira builds AI systems from Doha for businesses in Qatar, the GCC and worldwide.

Let’s build something people remember.

A rough idea is enough to start. Tell us the problem, and we will come back with a plan and a number.