Build an AI product with an independent developer

What an AI product is, and is not

A large language modeli is a building block, not a product. The product is the job around it: what goes in, what comes out, who checks it and how it fits into someone's day.

People build assistants that answer from company documents, tools that summarise calls, helpers that draft emails or listings, systems that sort and tag incoming messages, and extraction tools that turn invoices or forms into data. The frame around the model is usually a web app or a SaaS.

Typical building blocks

Find My Stack can suggest tools for your case. Common parts are these:

• A model provider through an APIi, such as OpenAI or Anthropic. See OpenAI developers.
• Promptsi that are written, tested and versioned like code.
• Retrieval augmented generationi when answers must come from your documents.
• A vector databasei such as the pgvector extension in Postgres.
• Evaluationsi to measure quality instead of guessing.
• Guardrails, usage limits and logging, so a single user cannot run up costs or see another person's data. Python and Node.js are the usual back end languages.

Models and their prices change often. Ask your developer to keep the model choice in one place of the code, so switching later is easy.

An example plan in milestones

Start with proof that the model can do the job. Each milestone is funded before work starts and runs 1, 2, 3, 4, 6 or 8 weeks. Milestones are from 50 to 2,000 each in your project currency.

1. For example, week 1 to 2, feasibility: collect twenty or so real examples, try models and prompts, report what works and what does not. You decide to continue with facts in hand.
2. For example, weeks 3 to 6, working product: the app around the model, accounts, saving results, a first retrieval setup if needed.
3. For example, weeks 7 to 8, quality and safety: evaluation set, handling of wrong or empty answers, rate limits, privacy review of what is sent to the provider.
4. For example, weeks 9 to 10, launch: monitoring of cost and errors, feedback buttons for users, handover.

The weeks are for illustration. Developers set the length of each milestone in their proposal, and each real milestone runs 1, 2, 3, 4, 6 or 8 weeks. Cut the first one short on purpose: its main output is a decision.

What belongs in the project description

• The task in one sentence and who benefits.
• Examples of good inputs and the answers you would be happy with. These become the test set.
• Your data: what exists, in which format and how sensitive it is.
• How wrong answers should be handled and who reviews critical results.
• Privacy needs, for example no personal data sent to outside providers.
• Expected number of users and how often they use the product, since model use is billed by volume.
• Whether you prefer a specific model or provider.

If personal data is involved, data protection rules may apply. This is general information, not legal or tax advice.

How to recognise a suitable developer

Be wary of anyone who promises perfect answers. A good developer talks about testing, fallbacks and limits, shows past work with models in real products and explains costs per request in simple terms. They should be happy to build the evaluation set with you.

On ErdiBuilds, each proposal carries a match scorei. A developer needs at least 40 to apply, and 80 or more gets a Top match badge. Look at the Expert in list for AI related technologies. Developers are independent professionals, not our employees.