Hokayantra

Getting the model to work inside what you already run

Calling a model API is a morning's work. Making that call behave correctly inside an existing codebase, with your permission model, your data residency rules and your on-call rota, is the actual project.

Scope

A model touches more of your stack than expected

It needs data, which means pipelines and a freshness guarantee. It needs to respect who is allowed to see what, which means your auth system rather than a parallel one. It needs to be observable, because when output quality drops somebody will want to know which change caused it.

Each of those is ordinary engineering. The difficulty is that AI features are often built by a small team off to one side, and the integration work only gets discovered at launch.

Separate systems resolving to a single path

Integration surface

What we connect

Data
Warehouses, document stores, internal wikis, ticketing and CRM
Identity
Your existing provider, so permissions are not reinvented
Serving
Hosted APIs, or open-weight models on your own infrastructure
Observability
Traces and metrics into the stack your team already watches
Freshness
Refresh pipelines with a stated guarantee, not a nightly guess

The one people underestimate

Retrieval will happily leak documents

Retrieval systems are very good at surfacing documents a particular user was never meant to see. If your index does not carry the same access rules as the source system, you have built a search engine that ignores your permission model.

This has to be designed in at indexing time. Filtering results afterwards looks like it works, right up until someone changes team and the index does not notice.

Questions

Frequently asked

Can the model run on our own infrastructure?
Yes. Where data residency or contract terms rule out a hosted API, open-weight models on your own infrastructure work. The trade is operational: you take on serving, scaling and upgrades, and quality at a given size trails the frontier hosted models.
How do you stop retrieval leaking restricted documents?
Access rules travel with the content into the index, so filtering is part of retrieval rather than a cleanup step afterwards. More work upfront, and the only version that survives someone changing team.
Do you work alongside our engineers?
Usually, and it goes better that way. Integration touches systems your team knows far better than we will, so the sensible split is us on the AI-specific parts and your engineers on what they own.

Already have a model that works in isolation?

Tell us what it needs to talk to. Integration problems are usually specific, and a short conversation normally identifies the hard part.