Skip to content
Nubent

/AI-native: AWS + Claude

Claude on Amazon Bedrock, built into your product and your AWS account

We bring Anthropic's Claude models into your product through Amazon Bedrock. We pick the model for each feature by cost and latency, apply prompt caching, tool use and structured outputs, and keep guardrails, evaluation and quotas under control. Everything runs in your AWS account, under your IAM policies and your data governance.

/Diagnosis

Signs it is time

The prototype won't survive production

You have a Claude demo that works, but you don't know how it behaves under real traffic, what it will cost per month or what happens when you hit a quota.

Security wants it all inside AWS

Your security team or your customers require inference to run through your AWS account, with IAM, CloudTrail and encryption, not through yet another external API.

The token bill keeps climbing

Every call resends the same long context, everything runs on the largest model, and nobody tracks cost per feature or per customer.

Outputs you can't keep in check

The model sometimes returns formats that break your code, makes up data or answers topics it shouldn't, and you have no objective way to measure it.

/What we deliver

Concrete deliverables, in your repository

  • Integration architecture

    A documented design for how your application calls Bedrock: Converse API, inference profiles, error handling, retries, timeouts and streaming.

  • Model selection matrix

    A comparison of the Claude models available on Bedrock for each use case, measured on quality, latency and cost against your own data.

  • Cost optimization

    Prompt caching on repeated prefixes, routing to smaller models where they do the job, batch inference for deferred workloads and per-feature budgets.

  • Tool use and structured outputs

    Tools defined with JSON schemas, response validation and a secure connection to your internal APIs and databases.

  • Guardrails and evaluation

    Amazon Bedrock Guardrails tuned to your domain and a versioned evaluation set that runs in CI before every prompt or model change.

  • Infrastructure and governance in Terraform

    Least-privilege IAM permissions, invocation logging, monitored quotas, cost alarms and all the code in your repository.

/How we work

Our process

  1. Use case assessment

    With read-only access, we review the product, the data involved, latency requirements and the current state of your AWS account.

  2. Evaluation and design

    We build an evaluation set from real examples, test the candidate Claude models and define the architecture and the budget.

  3. Build by pull request

    We implement the integration, guardrails, caching and Terraform infrastructure, with every change reviewed in your repository.

  4. Launch and tuning

    We go live with quality, cost and latency metrics, tune prompts and quotas with real data and stay on through operations.

Why use Claude through Amazon Bedrock

Amazon Bedrock offers Anthropic's Claude models as a managed service inside AWS. For a company already running on AWS, that means the same billing, the same IAM identities, the same CloudTrail logging and the same private networks as the rest of the platform. Inference is consumed like any other service in the account.

That integration makes the conversation with security and compliance much simpler. Permissions are defined per role, every call is audited, and you can restrict which models each account may use with organization policies. For many teams, that is the difference between a pilot stuck in the lab and a feature that reaches customers.

Choosing a model by cost, latency and quality

The Claude models on Bedrock have different profiles of capability, speed and price. Using the largest one for everything tends to be expensive and slow; using the smallest one for everything tends to fail on the hard tasks. The right call depends on each feature, and it should be made with your own data rather than generic benchmarks.

That is why we start with an evaluation set built from real examples from your product. We measure each candidate model on quality, latency and cost per request, and document the choice. When a new model ships, the same set lets you compare quickly and migrate with confidence instead of switching blind.

Prompt caching, tool use and structured outputs

Many applications resend the same long context on every call: system instructions, documentation or tool definitions. Bedrock supports prompt caching on the models that offer it, so that prefix can be reused, cutting cost and latency. We structure prompts so the stable part comes first and hits the cache.

With tool use, Claude can query your APIs, databases or internal services through tools we define with JSON schemas. Combined with structured outputs and validation in your code, the model stops returning free text that is hard to parse and starts producing data your application can use safely.

  • Stable prefixes at the start of the prompt to make the most of the cache.
  • Least-privilege tools with input validation.
  • Responses validated against a schema before they reach the user.

Guardrails, quotas and cross-region inference

Amazon Bedrock Guardrails filters denied topics, harmful content and sensitive data on both inputs and outputs. We configure it for your domain and test it with adversarial cases inside the evaluation set, so every prompt change is validated before it reaches production.

In production, token and requests-per-minute quotas matter too. We monitor them in CloudWatch, set up retries with backoff and, when it makes sense, use cross-region inference profiles to spread load across Regions. We check which profile fits where your data must be processed, so the technical decision respects your data residency requirements.

/FAQ

Claude on Amazon Bedrock: frequently asked questions

Another question? Write to hello@nubent.com and an engineer will answer.

What are the advantages of using Claude on Amazon Bedrock instead of the Anthropic API?

The main one is that inference stays inside your AWS account: it is billed with the rest of your services, controlled with IAM, audited with CloudTrail and integrated with your existing networks and policies. Anthropic's direct API is also a valid option. We choose based on where your platform lives, what your security team requires and which features you need.

Does Bedrock use my data to train models?

According to AWS documentation, Amazon Bedrock does not use your prompts or responses to train models and does not share them with model providers. Even so, we review with you what data travels in each call, enable invocation logging where appropriate, configure encryption and define retention, so your data governance is documented and auditable.

What is prompt caching and when should I use it?

Prompt caching lets Bedrock reuse a part of the prompt that repeats across calls, such as long instructions or reference documentation, instead of processing it again every time. It pays off when you have a large, stable prefix and a high volume of requests. We measure it in your case, because the savings depend on prompt structure and traffic patterns.

Which Claude model is right for my use case?

It depends on the task. Complex reasoning or coding usually justifies a more capable model, while classification, extraction or short answers tend to work well with faster, cheaper models. Instead of guessing, we build an evaluation set with your data and measure quality, latency and cost for each available model.

How do you keep Claude from making up answers in my product?

We combine several layers: context retrieved from your sources with RAG or tool use, clear instructions on what to do when there is no information, structured outputs validated by your code, and Amazon Bedrock Guardrails. Then we measure behavior with an evaluation set that includes unanswerable questions, to catch regressions before every deployment.

Next step

Let's talk about your Claude on Amazon Bedrock project

30 minutes with an engineer. Concrete next steps, whether or not you work with us.