Most places you can hire AI developers sell you a résumé and a rate. You interview for prompt fluency, onboard, and learn months later whether anyone has put a model in front of paying users.
We sell the output, not the seat. LLM features, retrieval and agents land in your branch with evals, spend limits and a rollback path already wired.
The distance between a demo and production is where AI work fails: no evaluation set, no cost ceiling, no plan for the day the model provider changes behaviour.
The work is LLM integration, retrieval over your own data, AI agents with a human gate on anything irreversible, and the workflow automation around them.
It is the same engagement as our staff augmentation, with one addition: nothing model-driven merges without an eval that shows what it does on your data.
The discipline comes from payments, where a wrong write has a cost. Anupam Shah ran payment infrastructure at $60M+ volume in MAS-regulated, SOC 2 / ISO 27001 environments.
Scope
Hosted and open-weight models; retrieval over the Postgres, vector store or search index you already operate; agents on a framework your team can maintain. AI developers for hire are usually sold by framework.
We scope by the workflow and the data, and pick the tooling second.
Before any model-driven feature ships, it is measured against examples from your own data: what good output looks like, what failure looks like, and what each call costs.
Prompt changes, model upgrades and retrieval changes all re-run the same set, so quality is a number rather than an impression.
The unit you are buying is finished, evaluated work. It lands in your repository with the reasoning written down, ready to deploy.
Everything is written inside your repository and assigned to you in the contract, prompts and eval sets included. Model provider accounts stay in your name.
The harness runs after the engagement ends, so a model upgrade six months from now can be judged on your data, in your CI.
Integration, retrieval, agents, evaluation and the automation around them. Training foundation models, data labelling at volume, and legal or regulatory advice sit with specialist firms.
Bring the workflow, the data and what you need built. The first call walks the use case and where it can fail; the scope follows in writing with a fixed fee, and work lands in your repository evaluated and tested.
A fixed fee against a written scope, never an hourly rate or a seat price. The scope sets the price, and it is written after the use case and the data have been walked with you.
To compare it with a hire, the hire-versus-engage model puts back the costs a salary comparison leaves out.
Ask what they have put in front of paying users, and how they knew it worked. An engineer who ships AI to production answers with an eval set, a cost per call and a rollback path.
A demo builder answers with a model name and a prompt.
Hire when the AI work is permanent and someone senior can direct it. Engage when it needs to be in production this quarter, or when you want to learn what the role is before hiring for it.
The full model prices both with your own numbers.
Yes. The work happens in your repository, your review process and your tooling. Your engineers can read every diff, run every eval and take the work over when the engagement ends; the handover is built for exactly that.
You do. Code, prompts, eval sets and configuration are written in your repository and assigned to you in the contract. Your data stays in your systems, and the model provider accounts stay in your name.
Retainers end on 30 days written notice from either side, with no minimum term. Milestone billing means you never pay ahead of delivered work.
The start date is set in the signed scope, and work begins as soon as the deposit clears.
Founder, Provingmark
Thirty minutes, no charge.
Related: Staff augmentation — the general page · LLM & RAG integration · AI agent development · Hire vs engage a firm: a cost model · What staff augmentation costs · How we scope and bill · About Anupam Shah