◇ Could this help me?
Yes on the cost reasoning, with one recommendation that needs checking against your own case. The service-by-service guidance is broadly accurate: on-demand NoSQL, object storage, serverless compute, and managed queues really do stay near zero at low volume, and calling a model provider directly is usually cheaper than going through a reseller layer. The part to verify before committing is packing an open source model into a serverless function, which runs into real limits on package size, cold start time, execution timeout, and the absence of a GPU, so it works for small models and not for everything. Also weigh what the managed model service offers beyond price, such as data residency and enterprise controls, before ruling it out on cost alone. This is a general practice.
Review this project's cloud spend by service and tell me where the money actually goes. For each pay-per-use service, tell me whether a cheaper architecture exists: on-demand pricing rather than provisioned capacity, scale-to-zero compute rather than always-on, calling a model provider directly rather than through a reseller layer, or self-hosting a small open source model rather than paying per call. For any self-hosting suggestion, state the package size, cold start, timeout, and hardware constraints honestly so I can judge whether it is realistic here.