Applied AI
AI features that survive contact with users.
Anyone can wire a model to a text box. The work is everything around it: what it is allowed to see, what it is allowed to do, how you know it improved, and what it costs when a thousand people use it at once.
The argument
“Can we not just ask ChatGPT?”
You can, and you should — for the one-off. The difference appears the moment the answer has to be repeatable, attributable and cheap at volume.
A chat window and a prompt
- Reads a slice of the context and sounds certain about the rest
- No record of what it looked at, so a wrong answer cannot be explained
- Every prompt edit is an untested deploy
- Cost per month is discovered at the end of the month
- Nothing stops it acting on data the user should not have seen
An Arnavia AI system
- Retrieves from the whole corpus, and cites the passage it used
- Model input logged, so every answer is traceable to its sources
- Evaluation set scores every change; a regression fails the build
- Cost per request measured in advance and capped per tenant
- Permissions resolved before retrieval, not after generation
What we build, in the order we build it
First the evaluation set. Before a prompt is written, we assemble real questions and real answers from your own history — the tickets, the emails, the decisions. Without it, every later conversation about quality is an opinion.
Then the retrieval. Chunking that respects the document’s structure, permission filters that run before the search, and hybrid scoring so exact identifiers work as well as prose.
Then the model, last. The smallest one that clears the score. Routing for the easy majority, a larger model for the hard tail, caching for the repeats, and a hard ceiling per tenant.
Then the interface. A confidence threshold, a refusal path that reads like a sentence, and a human in the loop wherever the action is irreversible.
- GroundingChunking, embeddings and retrieval tuned against real questions
- EvaluationA labelled set, a score, and a gate before every deploy
- GuardrailsSchema-constrained output, refusal paths, human review
- PrivacyYour documents stay in your infrastructure; no training on your data
- CostBudgets per tenant, caching, smaller models where they suffice
- AuditWhat the model was given, what it returned, and who was allowed it
What done looks like
Four artefacts, and then it is not our opinion any more.
A scored baseline
The model is measured against your current human process on the same test set. If it does not beat it, we say so before you spend the build.
A deploy gate
Any change that lowers the score fails in CI. That is the difference between an experiment and a feature.
A cost figure
Cost per request, per user and per month at your real volume — with a ceiling that holds.
An escape hatch
A documented path for the cases the model must not handle, so nothing important depends on a probability.
Next step
Bring us the workflow you have given up on.
Two weeks of review tells you whether it is worth automating, what it would cost per month to run, and what it would take to prove it works.