Evals before code
We build the golden dataset first. We know what "right" looks like before we run a prompt. Most teams skip this and pay for it in regressions the moment they change a model.
02 AI Product Development
AI features that hold up
when the novelty fades.
Strategy, RAG, agent systems, evaluation harnesses, prompt libraries, cost models. We will build the AI feature your roadmap calls for, or tell you why your roadmap is wrong, with the math to back it up.
A What we build
Seven ways we ship AI product work.
AI Product Strategy Sprint
Two-week discovery. Hypothesis, use-case shortlist, spec, cost model, go or no-go recommendation.
$14,000
RAG MVP
Retrieval + chat on your docs or data. Pipelines, retrieval quality evals, UX.
$32,000
Agent System MVP
Single multi-step use case end-to-end. Claude Agent SDK, tool guardrails, observability.
$55,000
AI Feature Integration
Copilot, summarize, classify, semantic search into an existing product.
$28k - $75k
Full AI MVP
End-to-end greenfield AI product. 12 to 16 week delivery.
$120k - $280k
Evaluation Harness Setup
Prompt versioning, regression suite, offline and online evals, dashboards.
$18,000
Product Retainer
Ongoing AI product engineering plus eval maintenance.
$14,000 / mo
B The parts most teams skip
Where AI products actually fall over.
Cost models that survive prod
We map request mix to models, cache ratios to tokens, and build a spreadsheet that tells you the cost at 10x and 100x. Prompt caching baked in from day one.
Prompt injection as a first-class threat
Our security team reviews every tool-use surface. Client data never flows into system prompts. We red-team before launch, not after the incident.
Fallback paths for bad days
Model outage, context overflow, user-provided ambiguity. Every AI feature ships with graceful degradation, not "it stopped working."
C What you get
Deliverables.
The product or feature
Shipped, observable, cost-forecasted, eval-gated.
Prompt & eval library
Versioned prompts, regression tests, few-shot examples, routing logic. Yours to own.
Your AI Product Handbook
Model routing, cost forecast, safety policy, observability map. Reads like a real playbook because it is.
Sample →D Further reading
The thinking behind the work.
Essays
-
What Claude tool use actually unlocks12 min / The mechanism that turns a prompt into a program.
-
Prompt caching and unit economics14 min / The cost model most teams never run.
-
Extended thinking as research budget9 min / When to pay for more tokens of thought.
Have an AI feature on the roadmap?
Send the idea. We will come back with a strategy sprint scope or a "do not build this" note with reasoning.