## Build with AWS AI



 

# AI solutions on AWS.Scoped, built, and deployed to production

QED42 builds on Amazon Bedrock and AWS AI services for enterprises that need AI in their existing systems, not as a standalone experiment

 [Share a use case](/contact) 

 

 

 

 



 ## Engagements

 

 

### AI Readiness Assessment

We review your AWS environment, data sources, and candidate use cases. The output is a ranked build plan: what can go to production now, what needs data or infrastructure work first.

 

### Data Foundation Engineering

Pipelines, vector storage, embedding strategy, access controls, and a governed data layer for the AI workloads that follow. For organisations whose data sits across disconnected systems or in formats foundation models cannot consume.

 

 

### Use Case Build

One defined use case, from architecture through production. Model selection on Bedrock, fine-tuning where the use case requires it, evaluation metrics, human review checkpoints, and handover documentation.

 

### Agentic AI Deployment

Multi-step agents that reason across tools, hold context, and execute workflows with guardrails. Policy enforcement, monitoring, evaluation loops, and escalation paths built into the deployment.

 

 

 

 

 

 



 ## AWS AI services we work with

 

 

QED42 works across the AWS AI stack — selecting the right services for the use case, not defaulting to the same stack for every client.

01

### Amazon Bedrock

Foundation model access with enterprise security. We use Bedrock for RAG implementations, knowledge bases, content workflows, and agent orchestration. We select from the full range of foundation models available on Bedrock, matching model capabilities to the use case rather than defaulting to one provider.

 

02

### Amazon Bedrock Agents

Production infrastructure for autonomous agents. Tool access, memory, session isolation, and governance for agents that operate across enterprise systems.

 

03

### Amazon Q

Enterprise knowledge assistant. We connect Q to your existing data sources, from content systems and business applications to databases and internal tools, so teams get answers from organisational data. Similar to what our product Aeldris does.

 

 

04

### Document and language processing

Textract for document extraction and Comprehend for entity recognition, sentiment analysis, and classification. These services feed into larger pipelines for intake automation, content routing, and compliance workflows.

 

05

### Amazon SageMaker AI

Infrastructure for fine-tuning and deploying models where Bedrock's managed options do not meet the use case requirements. We use SageMaker AI for custom training jobs, model evaluation, and hosting end points that need dedicated compute.

 

 

 

 

 

 



 ## Why QED42 for AWS AI

 

 

### We build the systems AI runs on

Content platforms, customer-facing applications, enterprise websites. QED42 has delivered platform engineering for 17 years. When we add AI to a system, it shows up where editors and end users already work.

 

### Fine-tuning, not model training

We select from the foundation models on Bedrock and fine-tune smaller LMs where a specific task demands higher accuracy. We do not train models from scratch. The value is in getting the right model to perform well on your data.

 

 

### Leading the Drupal AI initiative

Drupal powers content infrastructure for governments, universities, publishers, and large enterprises. QED42 leads the innovation workstream that defines how AI integrates with that ecosystem at the architecture level.

 

### Human review at every decision point

Every system we deploy has defined points where a person reviews, approves, or overrides AI output before it reaches an end user. That is the design principle, not a feature toggle.

 

 

 

 

 

 



## Have something to build on AWS?

 



 [ Talk to us  ](/contact) 

 

 

 

 

 

 

 FREE GUIDE

### Why most AI pilots never reach production

The gaps that stall AI projects between a working demo and a deployed system. Data quality, integration, monitoring, governance, and team adoption. Applies whether you are on Azure, AWS, or both.

 

 

 You must have JavaScript enabled to use this form.

 Name  

 Work email\*  



 Leave this field blank  



 



 

 

 

 

 



{"@context":"https://schema.org","@type":"FAQPage","@id":"https://qed42.com/ai-solutions/build-with-aws-ai#faq","isPartOf":{"@id":"https://qed42.com/ai-solutions/build-with-aws-ai#webpage"},"mainEntity":[{"@type":"Question","name":"What does it take to make enterprise data ready for Amazon Bedrock?","acceptedAnswer":{"@type":"Answer","text":"Enterprise data does not need to be fully reorganized before using Amazon Bedrock.The key is making the right data sources securely accessible for the intended usecase, whether through retrieval pipelines, search layers, APIs, or vector-based approaches, while preserving existing access controls and governance. Bedrock can work with structured data, unstructured documents, APIs, and existing enterprise applications. Only the data relevant to the target use case needs to be accessible,not everything. QED42 runs a two-week readiness assessment that maps which data sources are needed, how to connect them, and what preparation is required before the build begins."}},{"@type":"Question","name":"How do Amazon Bedrock costs scale from pilot to production?","acceptedAnswer":{"@type":"Answer","text":"Bedrock charges per token for on-demand inference, with significant price differences across models. The cost risk at scale comes from routing all tasks through a single large model without evaluating whether a smaller, more efficient model would perform equally well. In many cases, a tiered approach works: using lighter models for high-volume tasks and reserving larger models for complex work.But the right strategy depends on the data and the use case. QED42 architects the cost model during the build and deploys spend monitoring with alerts so there are no surprises when production traffic increases."}},{"@type":"Question","name":"When should we use RAG on Bedrock versus fine-tuning a model?","acceptedAnswer":{"@type":"Answer","text":"RAG (retrieval-augmented generation) works best when responses need to be grounded in specific, frequently updated organisational data like policy documents, product catalogues, or knowledge bases. The model retrieves relevant content from a vector store at query time and uses it to generate an answer. Fine-tuning is better when the model needs to consistently perform a specialised task, such as classifying domain-specific documents, where the base model's general knowledge falls short. RAG is faster to set up and does not require retraining when data changes. Fine-tuning produces a more focused model but takes longer to prepare and validate. QED42 evaluates both during the use case build and selects based on accuracy, latency, and maintenance needs."}},{"@type":"Question","name":"How does Amazon Bedrock handle data privacy and security?","acceptedAnswer":{"@type":"Answer","text":"Bedrock keeps all inputs and outputs within your AWS account boundary. Your data is not used to train the underlying foundation models. Access is controlled through IAM roles, so you can restrict exactly which services, users, and applications can invoke specific models. Bedrock supports VPC endpoints for private network access, encryption at rest and in transit, and is HIPAA eligible for healthcare workloads. Bedrock Guardrails adds a configurable layer that can filter harmful content, block restricted topics, and redact sensitive information from model inputs and outputs. QED42 configures these security and governance controls as part of every production deployment."}},{"@type":"Question","name":"How do you select the right foundation model on Bedrock?","acceptedAnswer":{"@type":"Answer","text":"Bedrock provides access to a range of foundation models, each with different strengths in reasoning, latency, cost, and task specialization. The right choice depends on what the use case demands: accuracy requirements, acceptable response time, per-request cost budget, and whether open-weight flexibility matters. QED42 benchmarks candidate models against your actual data during the build and documents the selection rationale so the decision can be revisited as new models become available on Bedrock."}},{"@type":"Question","name":"What does it take to deploy AI agents on AWS in production?","acceptedAnswer":{"@type":"Answer","text":"A production AI agent on AWS needs more than a working prototype. Bedrock Agents provides the orchestration layer, but a production deployment also requires session memory so the agent retains context, tool access with correct IAM permissions, evaluation frameworks that test agent behaviour against expected outcomes, and policy enforcement that defines what the agent can and cannot do. AgentCore, generally available since late 2025, adds runtime isolation, observability, and identity management. QED42 builds agents with all of these layers configured, including defined escalation paths for decisions that require human review."}},{"@type":"Question","name":"How long does it take to go from use case to production on AWS AI?","acceptedAnswer":{"@type":"Answer","text":"A readiness assessment that evaluates data, infrastructure, and use case fit takes two weeks. A focused use case build, from architecture through production deployment, typically runs eight to twelve weeks. Agentic AI deployments take longer depending on how many tools and systems the agent orchestrates. The biggest variable is data readiness. Organisations with clean, accessible data in S3 or structured databases move faster. Those with fragmented data across legacy systems may need a data foundation engagement before the use case build starts. QED42 agrees a fixed timeline with the client before every engagement begins."}}]} ## Common questions

 



  ### What does it take to make enterprise data ready for Amazon Bedrock?

    Enterprise data does not need to be fully reorganized before using Amazon Bedrock.The key is making the right data sources securely accessible for the intended usecase, whether through retrieval pipelines, search layers, APIs, or vector-based approaches, while preserving existing access controls and governance. Bedrock can work with structured data, unstructured documents, APIs, and existing enterprise applications. Only the data relevant to the target use case needs to be accessible,not everything. QED42 runs a two-week readiness assessment that maps which data sources are needed, how to connect them, and what preparation is required before the build begins.

 

  

  ### How do Amazon Bedrock costs scale from pilot to production?

    Bedrock charges per token for on-demand inference, with significant price differences across models. The cost risk at scale comes from routing all tasks through a single large model without evaluating whether a smaller, more efficient model would perform equally well. In many cases, a tiered approach works: using lighter models for high-volume tasks and reserving larger models for complex work.But the right strategy depends on the data and the use case. QED42 architects the cost model during the build and deploys spend monitoring with alerts so there are no surprises when production traffic increases.

 

  

  ### When should we use RAG on Bedrock versus fine-tuning a model?

    RAG (retrieval-augmented generation) works best when responses need to be grounded in specific, frequently updated organisational data like policy documents, product catalogues, or knowledge bases. The model retrieves relevant content from a vector store at query time and uses it to generate an answer. Fine-tuning is better when the model needs to consistently perform a specialised task, such as classifying domain-specific documents, where the base model's general knowledge falls short. RAG is faster to set up and does not require retraining when data changes. Fine-tuning produces a more focused model but takes longer to prepare and validate. QED42 evaluates both during the use case build and selects based on accuracy, latency, and maintenance needs.

 

  

  ### How does Amazon Bedrock handle data privacy and security?

    Bedrock keeps all inputs and outputs within your AWS account boundary. Your data is not used to train the underlying foundation models. Access is controlled through IAM roles, so you can restrict exactly which services, users, and applications can invoke specific models. Bedrock supports VPC endpoints for private network access, encryption at rest and in transit, and is HIPAA eligible for healthcare workloads. Bedrock Guardrails adds a configurable layer that can filter harmful content, block restricted topics, and redact sensitive information from model inputs and outputs. QED42 configures these security and governance controls as part of every production deployment.

 

  

  ### How do you select the right foundation model on Bedrock?

    Bedrock provides access to a range of foundation models, each with different strengths in reasoning, latency, cost, and task specialization. The right choice depends on what the use case demands: accuracy requirements, acceptable response time, per-request cost budget, and whether open-weight flexibility matters. QED42 benchmarks candidate models against your actual data during the build and documents the selection rationale so the decision can be revisited as new models become available on Bedrock.

 

  

  ### What does it take to deploy AI agents on AWS in production?

    A production AI agent on AWS needs more than a working prototype. Bedrock Agents provides the orchestration layer, but a production deployment also requires session memory so the agent retains context, tool access with correct IAM permissions, evaluation frameworks that test agent behaviour against expected outcomes, and policy enforcement that defines what the agent can and cannot do. AgentCore, generally available since late 2025, adds runtime isolation, observability, and identity management. QED42 builds agents with all of these layers configured, including defined escalation paths for decisions that require human review.

 

  

  ### How long does it take to go from use case to production on AWS AI?

    A readiness assessment that evaluates data, infrastructure, and use case fit takes two weeks. A focused use case build, from architecture through production deployment, typically runs eight to twelve weeks. Agentic AI deployments take longer depending on how many tools and systems the agent orchestrates. The biggest variable is data readiness. Organisations with clean, accessible data in S3 or structured databases move faster. Those with fragmented data across legacy systems may need a data foundation engagement before the use case build starts. QED42 agrees a fixed timeline with the client before every engagement begins.

 

  

 

 

 



 ### Bring your use case, we will build it

Not sure where AI fits? We will help you figure that out

 

 

 You must have JavaScript enabled to use this form.

 First name  

 Work email\*  



  What best describes where you are?   Exploring where AI fits 

  Have a defined use case 

  Ready to run a pilot 

  Scaling AI to production 



 

 Anything else you would like to share 

 



 [Privacy policy](/privacy-policy) 

 Leave this field blank