A model is not a product. The system around it is.
Red Sentra builds AI capabilities into software that has to keep working: the inference path, the services around it, and the rules that decide when the feature is allowed to speak at all.
Responsibility
What we take on.
- The AI capability inside a product: what it may claim, and what it must refuse
- The inference path, and what runs in the process serving it
- The product services either side of the model
- The grounding rules, and the record of when they were not met
- Operating the feature once real people depend on it
Engineering
The boundary is the feature
An assistant that answers confidently when it should not is worse than one that declines, because a product records what it was told. Vestry checks a set of grounding conditions before its assistant is allowed to speak, and failing any one of them means silence rather than a hedged answer.
It also keeps a record of the occasions it declined, which is what turns restraint into something a reader can audit instead of something the vendor asserts. An answer cites the source message it relied on, so the people who wrote the source material are the ones who can catch it being wrong.
What the product is allowed to write down
NutriSense reads a meal written in plain language and resolves it into identified foods with quantities. Nothing enters the diary until the numbers have been shown and confirmed, because a diary is a record and a wrong record is carried forward into every total after it.
The same rule separates an estimate from an observation: what a watch measured and what was eaten are kept as different facts, and one is never quietly promoted into the other. A photograph is analysed and then discarded rather than stored.
Choosing a model is a trade-off, not a preference
RustSTT runs four decoder architectures that are not variants of one design: a Conformer encoder with a CTC head, a FastConformer with a token-and-duration transducer, an audio encoder feeding a language-model decoder, and an encoder–decoder transformer. They differ in latency, in memory and in what they get right.
Holding all four behind a single decoder trait is what makes the choice a server-side decision rather than a commitment the product has to live with. A system that can only run the model it was built around has priced in an assumption that will expire.
Where an AI engagement starts
It can be one capability inside a product somebody else is building — an assistant, a retrieval layer, an inference service behind an existing interface — or the whole product with the AI part as its centre. Both are engagements; the difference is what the engineering responsibility is drawn around.
If the object is the product rather than the capability, custom software development is the closer description of the work, and the AI part is one of the things it contains.
Evidence
Systems this rests on.
Each entry names the contribution Red Sentra declared on that system. Nothing here is cited for work the content model does not carry.
NutriSense
A bounded assistant that turns a meal written in plain language into identified foods with quantities, shown for confirmation before anything is recorded.
Vestry
An assistant that answers only from a community’s own message history and cites the source message it relied on, with grounding conditions checked before it may speak.
RustSTT
Four decoder architectures held behind a single decoder trait, so which model serves a stream is a server-side decision rather than an integration one.
Limits
What this page does not claim.
- No accuracy, word-error-rate or benchmark figure is published for any model described here, and no model is claimed to be trained or owned by Red Sentra.
- No user count, request volume or hardware fleet size is stated. RustSTT runs on CUDA in production; nothing here says at what scale.
- No model-training, fine-tuning or data-labelling programme is claimed. The work described is engineering around models inside products.
- No customer is named as an AI reference, and no third-party model provider is described as endorsing Red Sentra.