Generative AI Development
Generation wired into a real product surface, arriving as a draft, with versioned prompts and a cost you can see.
Scope a surface ↗
What is generative AI development?
Generative AI development is the engineering around a model that writes something — a summary, a reply, marketing copy, a block of code, an image — so that it lands somewhere useful in your product rather than in a chat window. The interesting decisions are about the surface, not the prompt.
Output that arrives as a draft for a person to edit is a fundamentally different product from output that publishes itself, and the second needs far more evidence behind it. So we default to drafts, use schema-validated structured outputs, keep prompts in the repo with a regression suite, and put cost per accepted output on a dashboard — so quality and spend both stop being matters of opinion.
What a generative feature includes
Surface design
Where the output appears, what state it arrives in, and how someone accepts, edits or rejects it — decided before a prompt is written.
Versioned prompts
Prompts and context templates live in the repository with tests, so a change can be reviewed, released and rolled back like any other code.
Structured outputs
Schema-constrained responses with validation and repair, so downstream code isn’t parsing prose and hoping for the best.
Acceptance suite
A fixed set of inputs with expected characteristics, rerun on every prompt or model change to catch the quiet regressions.
Review and audit trail
Who generated what, from which inputs, on which model version — kept for the day somebody asks how a sentence got published.
Cost and rate control
Per-feature budgets, caching of repeated context, and hard limits so a retry loop or a launch can’t produce a surprise invoice.
How a generative feature ships
Pick one surface
We choose a single place in the product where generated output has obvious value, and write down what good enough looks like there.
Set the acceptance bar
A sample of real inputs, graded outputs, and a clear line between failures that are unacceptable and ones that are merely annoying.
Build the generation path
Context assembly, structured output, retries and fallbacks, behind an interface that doesn’t care which model is underneath it.
Put a person in the path
Drafts, edit tracking and a one-click reject, so the feature earns autonomy from measured accuracy rather than being handed it.
Measure use and cost
Acceptance rate, edit distance and cost per accepted output, reviewed together to decide whether to widen the surface or pull it.
Generative AI Development FAQ
A first surface is usually $10k to $35k to build. Running it is the part people underestimate, so we model token cost per accepted output during scoping — occasionally that maths alone kills the idea, which is cheaper than finding out in month three.
The tools we build with
Prompts are treated as code and outputs as drafts, which is what makes the model underneath a replaceable detail rather than an architectural commitment.
Generation
Anthropic
OpenAI
Mistral
OllamaProduct surface
Next.js
React
TypeScript
Node.jsOrchestration
Python
FastAPI
LangChain
RedisAudit & cost
PostgreSQL
OpenTelemetry
Datadog
SentryRelated work
Related reading
Know where generation would earn its place?
Describe the surface and what a good output looks like. We’ll come back with a scoped build, an estimated cost per accepted output, and the review flow around it.