AI integration into existing products
The interesting version of this work is never the model call. It is everything around it: what you retrieve before you ask, how you know the answer got worse after a change, what happens when the provider is down, and whether the feature justifies its bill. Those are backend problems, and they are the ones that decide whether the feature survives contact with real users.
We have shipped retrieval and classification inside products. We have not shipped a standalone AI product.
The moderation pipeline in Amour is ours, built around AWS Rekognition and Comprehend, and so is the retrieval layer behind an internal assistant for a client we cannot name. If you need a team whose whole business is models, that is not us, and we would rather say it here than in the second meeting.
Four ways this arrives.
A prompt that behaved on twenty hand-picked inputs meets ten thousand real ones. We build the evaluation set from your actual traffic first, so a change can be judged rather than argued about.
Usually a large context sent on every call, or a big model doing a job a small one can do. We measure where the tokens go before recommending anything.
Logging that captures the prompt, the retrieved context, the model version and the output, so a support question has an answer that is not a guess.
A fallback path, a timeout that is shorter than your user’s patience, and a degraded mode that is honest about being degraded.
Asked before every AI project.
Will you tell us if we do not need a model?+
Yes, and it happens often. Several requests we have taken turned out to be a search problem, a rules engine, or a form with better defaults. We would rather say that in week one than bill for a year of prompt tuning.
Whose API keys?+
Yours, in your accounts, with the spend visible to you. We never proxy your traffic through infrastructure we control.
What about our data going to a provider?+
We map exactly what leaves your systems and where it lands before anything is wired up. Where that is not acceptable, self-hosted open-weight models on your own hardware are a real option and we will price both.
Do you fine-tune?+
Rarely, and not as a first move. Retrieval and prompt structure fix most of what people bring to us as a fine-tuning problem, at a fraction of the cost and with none of the retraining treadmill.