OpenAI integrations
GPT models integrated with structured outputs, tool use, streaming and hard guardrails.
- Engagement
- Fixed-scope engineering
- Typical timeline
- 4–5 weeks
- Starts with
- Twenty of the worst real examples you have, the blurry ones, the rotated ones, the ones your team complains about.
Your customers photograph documents on a phone and email them in, badly lit, at an angle, sometimes upside down. Someone reads each one, retypes four fields and files the image, and reading is the part a vision model handles well while retyping is the part nobody should still be doing.
Production-grade features rather than prompt experiments.
What we build
Concrete artefacts, handed over and documented.
- 01
Structured outputs, so the model returns fields your code can rely on instead of a paragraph you have to parse
- 02
A vision pipeline for photographs and scans, with a confidence score per field and the image crop shown beside each value for review
- 03
Batch processing for anything not needed immediately, at a fraction of the live price
- 04
A latency budget per call, with streaming where a person is waiting and none at all where a job runs overnight
What changes
Documents arrive as data rather than as attachments someone has to open
The overnight half of the workload costs materially less than the live half
Response time becomes a design decision instead of whatever the API happens to return
How it runs
01Pick the capability and the bar
Vision, voice, reasoning or generation, one of them, with the accuracy or latency target it has to hit written down.
02Build and evaluate
The pipeline, scored against your worst real examples rather than the clean ones.
03Integrate and tune
Wired into the product, with the batch and live paths separated so cost follows urgency.
Chosen per project. Named here so you can see the shape of it.
- OpenAI
- Vercel AI SDK
- Braintrust
- Postgres
Questions we get asked
Tell us the outcome, not the tooling.
Send us the situation you are in. We will tell you which discipline it belongs to, what we would do first and what it costs, including when the answer is to wait.