MODULE_03
AI integration and LLM features
Features that go into production, not demos that go into a deck.
- LLM integration
- Gemini
- On-device ML
- Vision
- Evaluation
- Guardrails
- Cost modelling
What gets built
Model-backed features inside existing products — extraction, classification, summarisation, assisted search, conversational surfaces — plus on-device ML and vision pipelines where the work has to happen on the handset rather than a server.
Integrations are built against whichever model fits the job and the budget, with the provider isolated behind an interface so switching one out later is a configuration change rather than a rewrite.
Evaluation is the deliverable
An AI feature without an evaluation set is a feature nobody can tell has regressed. We build the eval alongside the feature: a fixed set of inputs, the expected behaviour, and a score that runs in CI like any other test.
This is the same argument as the regression suite above, applied to a component that is non-deterministic and therefore needs it more, not less.
Guardrails, and saying no
Input validation, output constraints, refusal handling, cost ceilings and a fallback for when the provider is down — which it will be.
And an honest answer about whether the problem wants a model at all. A surprising amount of what gets specified as an AI feature is a database query with better copy, and it is cheaper, faster and more reliable built that way.
What to expect from an AI app development agency
Most of what is sold as AI app development in Nepal and everywhere else is a wrapper around someone else's API with a chat box on top. That is sometimes the correct build. It is worth knowing that is what you are buying, and worth paying accordingly.
The work that is actually worth contracting out is the part around the model: deciding what goes on-device and what goes to a provider, an evaluation set that catches a regression before a user does, fallbacks for timeouts and outages, and a cost model that does not surprise you in month three when usage triples.
Glow is the working example on this site — a face scan where the cheap deterministic checks run on the handset first and only a frame worth paying for reaches Gemini, with a manual path behind it so a failed model call never costs a booking. The case study says how it is put together.