AWS Serverless API Deployment
Lambda behind API Gateway, with the database connections and the failure paths thought through.
From $549 3–7 days
Amazon Bedrock / AI Infrastructure
Model access wired up properly — private networking, quotas, logging and a spend ceiling.
Price and scope
From $749
Typically $749–$2,199, fixed in writing before anything starts.
What moves it up
- Private networking, so prompts never traverse the public internet
- Quota increases, which are a request AWS can refuse and which take real lead time
- Retrieval over your own documents, which is a data pipeline attached to a model rather than a model call
What actually goes wrong
AI features get built against a model endpoint the way they would against any other API, and two things about this one are different. The first is that the cost is per token and unbounded by default — a retry loop, a prompt that grew, or one enthusiastic user can turn a predictable monthly figure into a four-figure day, and nothing stops it because nothing was told to. The second is that quotas are per-account and per-model, and the request to raise them is not instant: teams discover their limit on launch day, which is the worst possible day to start a support case.
How I find it
aws bedrock list-foundation-models --by-inference-type ON_DEMAND, per region — availability differsaws service-quotas list-service-quotas --service-code bedrock, against projected peak throughputaws bedrock get-model-invocation-logging-configuration — whether anything records what was sentA budget and an anomaly alert existing at all, which is usually the finding
What you get
- Model access through VPC endpoints, so requests do not leave your network
- IAM scoped per model and per application, rather than one role that can invoke everything
- Invocation logging to a destination you control, with a decision recorded about what is retained
- Budgets, anomaly detection and a hard alerting ceiling, configured before the first production request
- Quota headroom requested and confirmed ahead of launch, not on the day
Questions
Does my data get used to train the models?
Not on Bedrock — inputs and outputs are not used to train the underlying foundation models and are not shared with model providers. What you do have to decide is your own invocation logging: whether you keep prompts, where, and for how long. That is your choice and it gets made deliberately rather than by default.
How do I stop this costing an unpredictable amount?
Budgets and anomaly alerts first, then throughput limits at the application layer, then the model choice itself. Per-token pricing with no ceiling is the specific risk here, and it is addressed before the first production request rather than after the first surprising invoice.
Bedrock, or calling a provider API directly?
Bedrock if you want it inside your VPC, under your IAM, on your existing bill, with several models behind one interface. A direct provider API if you need a model or a feature Bedrock does not carry yet. It is a real trade and the answer is not always Bedrock.
Can you build the retrieval pipeline over our documents too?
Yes, and be aware that is the larger half of the work — ingestion, chunking, embeddings, a vector store and the access controls that stop one user retrieving another's documents. The model call is the easy part; the data pipeline around it is where the effort and the risk are.
If this is not quite it
AWS VPC & Network Architecture
Subnets, routing and security groups laid out so the next person can read them.
From $349 2–4 days
Want this done?
Tell me what you are running and I will come back with a fixed price and a date. If it turns out you do not need this, I will say that instead.
Prefer to talk? Book a free call ↗ · Or hire me on Upwork ↗ · Typical reply within one business day.
Before you enquire
When I answer
Sunday to Thursday, 09:00–18:00 EET. Outside that I will still look, but I will not promise a time.
No 24/7 desk, and I will not pretend otherwise
One person, one time zone. If round-the-clock cover is what you need, you need a team, and I will say so rather than sell you a plan that cannot deliver it.
Your AWS bill stays yours
You pay Amazon directly and you keep control of the account. Nothing here resells your infrastructure or sits between you and your own billing.
A price that starts with "from" is a starting price
Every service page lists exactly what pushes a quote above it, before you ask. You get a fixed number in writing before any work begins.