AI FEATURES

AI features that earn their invoice.

Model calls are the easy part. The cost curve, the failure modes and the guardrails are the work — and they are what we scope first.

01 THE WORK

What this covers

Every AI feature has two numbers attached to it: what it does for the user, and what it costs per user per month. Most projects only work out the first one, then discover the second on an invoice. We put a rupee figure against a feature during scoping, model the cost at ten times current usage, and build the caps and fallbacks before the feature ships.

The second thing that goes wrong is failure modes. A model that returns nothing, returns nonsense, or returns slowly is not an edge case — it is a Tuesday. Every AI path we build has a defined behaviour for all three, and a non-AI route for the user who hits them.

We also use these tools on our own work: phase-gated builds driven by a CLAUDE.md spec are how the studio ships. What we recommend is what we run.

02 WHAT IS INCLUDED

Four things every build of this kind carries.

01

Scoped to a measurable job

One user outcome, one metric. Features that cannot be measured do not get built.

02

Costed before it is built

Token maths in rupees, at current and 10× usage, with the four levers that cut the bill identified up front.

03

Guardrails, not hopes

Rate limits, spend caps, retries, timeouts and a defined fallback for every failure mode.

04

Evaluated, not vibed

A test set for the prompt, so a model or prompt change is a measured decision instead of a guess.

03 PROOF

Shipped, and still running.

  1. 01

    Cortix

    Claude driving real-time matchmaking logic inside a Flutter product.

  2. 02

    What AI actually costs

    The token maths we run before quoting an AI feature, in rupees.

  3. 03

    Your vibe-coded app in production

    The seven things that quietly break when AI wrote the code and it is already live.

04 FAQ

The questions that come up first.

What does an AI feature cost to run?

It depends entirely on tokens per interaction and interactions per user. We quote a per-user-per-month figure during scoping and model it at ten times current usage, because that is the number that decides whether the feature is viable, not the build cost.

Can you add AI to a product we already have?

Yes — this is most of the AI work we do. It usually starts as a two-week Product Sprint against one measurable outcome rather than a platform-wide project.

Which models do you use?

Mostly the Claude API, chosen per task rather than per project — a cheap fast model for classification, a stronger one for reasoning. The routing is part of the cost work.

Pricing, phases and payment terms for all three engagement models are on the services page.

Also from the studio: web development · mobile apps · e-commerce

NEXT STEP

Have a product to ship?

Start a project ↗ hello@napdesigns.com