AWS Partner Network (APN) Blog
JA Canada launches generative AI with Qalius and Amazon Bedrock
By: Benedikt Langens, Principal Solutions Architect – Qalius Consulting Inc.
By: Sal Bayat, Enterprise Architect – Qalius Consulting Inc.
By: Grace Kunjumon, Partner Solutions Architect – AWS
By: Anisha Batra, Solutions Architect – AWS
A small team with no dedicated AI group can launch a retrieval-augmented generation (RAG) AI chatbot into production on a short timeline. Junior Achievement Canada (JA Canada) proved it, partnering with Qalius Consulting Inc. to launch a generative-AI application in four months and achieve an immediate increase in website engagement. The Canadian non-profit wanted to unlock the information in their educational resources without staffing a large internal AI team or committing to a multi-year transformation program. The managed-AI capabilities of Amazon Bedrock, combined with a partner like Qalius, a Canadian AWS Advanced Tier Services Partner, are what made it possible for JA Canada.
To make their educational resources more accessible, JA Canada focused on creating a conversational experience for their website and learning-management system. Thanks to the managed-access model of Amazon Bedrock, connecting to an LLM was the straightforward part. The application also needed to fit into existing web experiences, support repeatable evaluation, and include practical controls for production use. It also had to ground responses in JA Canada content and curated external sources. Amazon Bedrock sped up development of the RAG index, safety evaluation, and safety guardrails. This let Qalius focus custom development on JA Canada’s unique data sources and web applications.
This post shares the four key ingredients in Qalius’ recipe for generative AI success.
Start with one bounded business outcome
Generative AI projects can grow quickly when teams begin with model selection or a long list of use cases. JA Canada and Qalius started with a specific goal: help visitors find relevant information about JA Canada resources.
That narrower goal shaped the technical design. Instead of training a model on JA Canada content, the team used retrieval-augmented generation (RAG) with Amazon Bedrock Knowledge Bases. JA Canada sources and curated external information were kept in separate knowledge bases. Internal JA Canada material was prioritized as the primary source, while the external knowledge base was used to supplement responses when more information was needed.
The design also allowed source material to be refreshed as information changed. Custom integrations connected JA Canada’s existing website and learning-platform experiences to the AWS-hosted application. You start with one use case to reduce the number of decisions the team has to make at the same time – a best practice regardless of organization size. It makes it easier to decide which content belongs in the system, what a good answer should include, and what the application must do before it’s ready for launch.
Add evaluation to the delivery plan
A prototype can look promising after a few successful conversations. A production application needs a more reliable way to check how it behaves across expected, unexpected, and inappropriate prompts.
JA Canada and Qalius treated evaluation as its own phase of the delivery plan. The team defined test questions with expected answers, including cases where source information was available, cases where it was unavailable, inappropriate requests, and adversarial inputs. The team used RAG evaluation in Amazon Bedrock Evaluations for automated assessment, paired with manual user acceptance testing through the application itself.
The loop was straightforward: test, review the results, adjust configuration, and test again. This loop gave the team a consistent way to judge whether changes were improving the application. This approach is especially useful for small teams. Evidence is generated from repeatable tests instead of relying on impressions from a handful of prompts.
Include safeguards and cost controls in launch requirements
Generative AI applications introduce operating concerns that are often deferred until late in the project. JA Canada and Qalius included safeguards and cost controls in the earliest stages of design.
Amazon Bedrock Guardrails provided configurable safeguards for evaluating user inputs and model responses against policies appropriate to the use case. The design also limited how often the application could be called, using AWS WAF IP-based rate limiting at the edge and Amazon API Gateway usage plans on the backend. To limit context size, the application kept only recent conversation history. This let the team control token consumption while still allowing visitors to ask follow-up questions.
The team included AWS service cost reviews after launch. Early cost estimates for generative AI workloads depend on assumptions about usage and retrieval activity. They also depend on how much source material there is and how often it changes. Once real users start interacting with the application, data on actual token consumption should be used to balance performance and cost. From the beginning, JA Canada and Qalius treated safety, cost, and application behavior as part of the implementation scope.
Reduce custom work with managed capabilities
While the project required custom engineering, Qalius used AWS services to reduce costs and speed up delivery.
Qalius built integrations for JA Canada’s WordPress and Moodle experiences. The team also developed source-ingestion workflows and application-specific conversation handling. However, for the underlying generative AI capabilities, the team used managed services from Amazon Bedrock. Amazon Bedrock provided building blocks for RAG, evaluation, and safeguards. This allowed Qalius to focus custom development on the parts specific to JA Canada. Managed services still require architectural integration, testing, and governance. However, they reduce the amount of infrastructure work that must be built before the user-facing application can be delivered. For a small project team, less time spent building generic infrastructure means more time spent on integrations, content flows, and application behavior.
Increased engagement
JA Canada saw a significant increase in the time engaged users spent on its website after enabling chat. In the quarter after launch, time on site rose 64% compared with the same quarter a year earlier. JA Canada attributes this increase to the richer experiences that an accessible knowledge base makes possible. The chat interface delivered these gains just by being available, with no additional promotion.
Conclusion
For organizations that want to get started with generative AI, the main takeaway is the timeline. Using Amazon Bedrock and Qalius’ approach to generative AI delivery, JA Canada and Qalius moved an application to live production in four months, and engagement increased right away. Key ingredients:
- Start with one bounded business outcome – Keep the first use case narrow. Look for places where better access to trusted information has clear value, then identify the information the application should use.
- Build evaluation into the delivery plan – Decide how quality will be measured and create expected-answer test cases before launch and repeat evaluation as the system changes.
- Include safeguards and cost controls – Include safeguards, usage limits, cost considerations, and deployment planning in the implementation scope. Review user feedback and AWS service costs soon after launch, then tune based on what you observe.
- Use managed capabilities – Use managed capabilities to reduce unnecessary engineering effort.
You don’t need unlimited resources to get started. Choose one clear use case, decide how you’ll measure answer quality, add practical controls, and save custom engineering for the parts that matter most. With those steps in place, you can take your own generative AI project to production.
Ready to get started? Explore the Amazon Bedrock Knowledge Bases documentation to build your first RAG application or contact Qalius to discuss how they can help your organization move from idea to production.
Related APN Blog posts:
- Building Next-Gen AI Education through GoML and Amazon Bedrock – an APN partner case study on building a RAG-based education solution with Amazon Bedrock Knowledge Bases and Guardrails.
- Building a privacy preserving chatbot with Amazon Bedrock – an APN post on building a chatbot with Amazon Bedrock, focused on privacy and safeguards
Qalius – AWS partner spotlight
Qalius Consulting Inc. is a Canadian AWS Advanced Tier Services Partner that develops cloud-native web applications, modernizes existing applications, and provides ongoing support and optimization for AWS environments.


