How Hexagon constructed an AI assistant utilizing AWS generative AI companies

This publish was co-written with Julio P. Roque Hexagon ALI.

Recognizing the transformative advantages of generative AI for enterprises, we at Hexagon’s Asset Lifecycle Intelligence division sought to reinforce how customers work together with our Enterprise Asset Administration (EAM) merchandise. Understanding these benefits, we partnered with AWS to embark on a journey to develop HxGN Alix, an AI-powered digital employee utilizing AWS generative AI companies. This weblog publish explores the technique, improvement, and implementation of HxGN Alix, demonstrating how a tailor-made AI resolution can drive effectivity and improve person satisfaction.

Forming a generative AI technique: Safety, accuracy, and sustainability

Our journey to construct HxGN Alix was guided by a strategic strategy centered on buyer wants, enterprise necessities, and technological issues. On this part, we describe the important thing parts of our technique.

Understanding client generative AI and enterprise generative AI

Generative AI serves numerous functions, with client and enterprise functions differing in scope and focus. Client generative AI instruments are designed for broad accessibility, enabling customers to carry out on a regular basis duties resembling drafting content material, producing photos, or answering basic inquiries. In distinction, enterprise generative AI is tailor-made to deal with particular enterprise challenges, together with scalability, safety, and seamless integration with present workflows. These techniques usually combine with enterprise infrastructures, prioritize information privateness, and use proprietary datasets to offer relevance and accuracy. This customization permits companies to optimize operations, improve decision-making, and preserve management over their mental property.

Business in comparison with open supply LLMs

We used a number of analysis standards, as illustrated within the following determine, to find out whether or not to make use of a industrial or open supply giant language mannequin (LLM).

The analysis standards are as follows:

• Price administration – Assist keep away from unpredictable bills related to LLMs.
• Customization – Tailor the mannequin to know domain-specific terminology and context.
• Mental property and licensing – Keep management over information utilization and compliance.
• Information privateness – Uphold strict confidentiality and adherence to safety necessities.
• Management over the mannequin lifecycle – By utilizing open supply LLMs, we’re capable of management the lifecycle of mannequin customizations based mostly on enterprise wants. This management makes certain updates, enhancements, and upkeep of the mannequin are aligned with evolving enterprise targets with out dependency on third-party suppliers.

The trail to the enterprise generative AI: Crawl, stroll, run

By adopting a phased strategy (as proven within the following determine), we had been capable of handle improvement successfully. As a result of the expertise is new, it was paramount to rigorously construct the proper basis for adoption of generative AI throughout totally different enterprise models.

The phases of the strategy are:

• Crawl – Set up foundational infrastructure with a deal with information privateness and safety. This section centered on establishing a safe and compliant basis to allow the accountable adoption of generative AI. Key priorities included implementing guardrails round safety, compliance, and information privateness, ensuring that buyer and enterprise information remained protected inside well-defined entry controls. Moreover, we centered on capability administration and price governance, ensuring that AI workloads operated effectively whereas sustaining monetary predictability. This section was crucial in establishing the mandatory insurance policies, monitoring mechanisms, and architectural patterns to help long-term scalability.
• Stroll – Combine customer-specific information to reinforce relevance whereas sustaining tenant-level safety. With a strong basis in place, we transitioned from proof of idea to production-grade implementations. This section was characterised by deepening our technical experience, refining operational processes, and gaining real-world expertise with generative AI fashions. As we built-in domain-specific information to enhance relevance and usefulness, we continued to bolster tenant-level safety to offer correct information segregation. The purpose of this section was to validate AI-driven options in real-world eventualities, iterating on workflows, accuracy, and optimizing efficiency for manufacturing deployment.
• Run – Develop high-value use circumstances tailor-made to buyer wants, enhancing productiveness and decision-making. Utilizing the foundations established within the stroll section, we moved towards scaling improvement throughout a number of groups in a structured and repeatable method. By standardizing finest practices and improvement frameworks, we enabled totally different merchandise to undertake AI capabilities effectively. At this stage, we centered on delivering high-value use circumstances that immediately enhanced buyer productiveness, decision-making, and operational effectivity.

Figuring out the proper use case: Digital employee

A crucial a part of our technique was figuring out a use case that will supply one of the best return on funding (ROI), depicted within the following determine. We pinpointed the event of a digital employee as an optimum use case due to its potential to:

• Improve productiveness – Recognizing that the productiveness of any AI resolution lies in a digital employee able to dealing with superior and nuanced domain-specific duties
• Enhance effectivity – Automate routine duties and streamline workflows
• Improve person expertise – Present speedy, correct responses to person inquiries
• Assist excessive safety environments – Function inside stringent safety parameters required by purchasers

By specializing in a digital employee, we aimed to ship important worth to each inside groups and end-users.

Introducing Alix: A digital employee for asset lifecycle intelligence

HxGN Alix is our AI-powered chat assistant designed to behave as a digital employee to revolutionize person interplay with EAM merchandise. Developed to function securely inside high-security environments, HxGN Alix serves a number of features:

• Streamline data entry – Present customers with fast, correct solutions, assuaging the necessity to navigate intensive PDF manuals
• Improve inside workflows – Help Buyer Success managers and Buyer Assist groups with environment friendly data retrieval
• Enhance buyer satisfaction – Provide EAM end-users an intuitive device to interact with, thereby elevating their general expertise

By delivering a tailor-made, AI-driven strategy, HxGN Alix addresses particular challenges confronted by our purchasers, reworking the person expertise whereas upholding stringent safety requirements.

Understanding system must information expertise choice

Earlier than choosing the suitable expertise stack for HxGN Alix, we first recognized the high-level system parts and expectations of our AI assistant infrastructure. By means of this course of, we made certain that we understood the core parts required to construct a strong and scalable resolution. The next determine illustrates the core parts that we recognized.

The non-functional necessities are:

• Regional failover – Keep system resilience by offering the power to fail over seamlessly in case of Regional outages, sustaining service availability.
• Mannequin lifecycle administration – Set up a dependable mechanism for customizing and deploying machine studying fashions.
• LLM internet hosting – Host the AI fashions in an surroundings that gives stability, scalability, and adheres to our high-security necessities.
• Multilingual capabilities – Be sure that the assistant can talk successfully in a number of languages to cater to our numerous person base.
• Security instruments – Incorporate safeguards to advertise protected and accountable AI use, notably with regard to information safety and person interactions.
• Information storage – Present safe storage options for managing product documentation and person information, adhering to business safety requirements.
• Retrieval Augmented Era (RAG) – Improve the assistant’s potential to retrieve related data from saved paperwork, thereby bettering response accuracy and offering grounded solutions.
• Textual content embeddings – Use textual content embeddings to characterize and retrieve related information, ensuring that high-accuracy retrieval duties are effectively managed.

Choosing the proper expertise stack

To develop HxGN Alix, we chosen a mixture of AWS generative AI companies and complementary applied sciences, specializing in scalability, customization, and safety. We finalized the next structure to serve our technical wants.

The AWS companies embrace:

• Amazon Elastic Kubernetes Service (Amazon EKS) – We used Amazon EKS for compute and mannequin deployment. It facilitates environment friendly deployment and administration of Alix’s fashions, offering excessive availability and scalability. We had been in a position to make use of our present EKS cluster, which already had the required security, manageability, and integration with our DevOps surroundings. This allowed for seamless integration and used present investments in infrastructure and tooling.
• Amazon Elastic Compute Cloud (Amazon EC2) G6e situations – AWS gives complete, safe, and cost-effective AI infrastructure. We chosen G6e.48xlarge situations powered by NVIDIA L40S GPUs—essentially the most cost-efficient GPU situations for deploying generative AI fashions below 12 billion parameters.
• Mistral NeMo – We selected Mistral NeMo, a 12-billion parameter open supply LLM in-built collaboration with NVIDIA and launched below the Apache 2.0 license. Mistral NeMo presents a big context window of as much as 128,000 tokens and is designed for world, multilingual functions. It’s optimized for operate calling and performs strongly in a number of languages, together with English, French, German, Spanish, Italian, Portuguese, Chinese language, Japanese, Korean, Arabic, and Hindi. The mannequin’s multilingual capabilities and optimization for operate calling aligned nicely with our wants.
• Amazon Bedrock Guardrails – Amazon Bedrock Guardrails gives a complete framework for implementing security and compliance inside AI functions. It permits the customization of filtering insurance policies, ensuring that AI-generated responses align with organizational requirements and regulatory necessities. With built-in capabilities to detect and mitigate dangerous content material, Amazon Bedrock Guardrails enhances person belief and security whereas sustaining excessive efficiency in AI deployments. This service permits us to outline content material moderation guidelines, limit delicate matters, and set up enterprise-level safety for generative AI interactions.
• Amazon Easy Storage Service (Amazon S3) – Amazon S3 gives safe storage for managing product documentation and person information, adhering to business safety requirements.
• Amazon Bedrock Information Bases – Amazon Bedrock Information Bases enhances Alix’s potential to retrieve related data from saved paperwork, bettering response accuracy. This service stood out as a managed RAG resolution, dealing with the heavy lifting and enabling us to experiment with totally different methods and clear up complicated challenges effectively. Extra on that is mentioned within the improvement journey.
• Amazon Bedrock – We used Amazon Bedrock as a fallback resolution to deal with Regional failures. Within the occasion of zonal or outages, the system can fall again to the Mistral 7B mannequin utilizing Amazon Bedrock multi- Area endpoints, sustaining uninterrupted service.
• Amazon Bedrock Immediate Administration – This function of Amazon Bedrock simplifies the creation, analysis, versioning, and sharing of prompts inside the engineering crew to get one of the best responses from basis fashions (FMs) for our use circumstances.

The event journey

We launched into the event of HxGN Alix by a structured, phased strategy.

The proof of idea

We initiated the venture by making a proof of idea to validate the feasibility of an AI assistant tailor-made for safe environments. Though the business has seen numerous AI assistants, the first purpose of the proof of idea was to be sure that we might develop an answer whereas adhering to our excessive safety requirements, which required full management over the manageability of the answer.

In the course of the proof of idea, we scoped the venture to make use of an off-the-shelf NeMo mannequin deployed on our present EKS cluster with out integrating inside data bases. This strategy helped us confirm the power to combine the answer with present merchandise, management prices, present scalability, and preserve safety—minimizing the chance of late-stage discoveries.

After releasing the proof of idea to a small set of inside customers, we recognized a wholesome backlog of labor gadgets that wanted to go reside, together with enhancements in safety, architectural enhancements, community topology changes, immediate administration, and product integration.

Safety enhancements

To stick to the stringent safety necessities of our clients, we used the safe infrastructure offered by AWS. With fashions deployed in our present manufacturing EKS surroundings, we had been in a position to make use of present tooling for safety and monitoring. Moreover, we used remoted non-public subnets to be sure that code interacting with fashions wasn’t related to the web, additional enhancing data safety for customers.

As a result of person interactions are in free-text format and customers would possibly enter content material together with personally identifiable data (PII), it was crucial to not retailer any person interactions in any format. This strategy offered full confidentiality of AI use, adhering to strict information privateness requirements.

Adjusting response accuracy

In the course of the proof of idea, it grew to become clear that integrating the digital employee with our merchandise was important. Base fashions had restricted data of our merchandise and sometimes produced hallucinations. We had to decide on between pretraining the mannequin with inside documentation or implementing RAG. RAG grew to become the apparent selection for the next causes:

•  We had been within the early levels of improvement and didn’t have sufficient information to pre-train our fashions
• RAG helps floor the mannequin’s responses in correct context by retrieving related data, decreasing hallucinations

Implementing a RAG system offered its personal challenges and required experimentation. Key challenges are depicted within the following determine.

These challenges embrace:

• Destruction of context when chunking paperwork – Step one in RAG is to chunk paperwork to rework them into vectors for significant textual content illustration. Nevertheless, making use of this methodology to tables or complicated constructions dangers dropping relational information, which may end up in crucial data not being retrieved, inflicting the LLM to offer inaccurate solutions. We evaluated numerous methods to protect context throughout chunking, verifying that vital relationships inside the information had been maintained. To deal with this, we used the hierarchical chunking functionality of Amazon Bedrock Information Bases, which helped us protect the context within the remaining chunk.
• Dealing with paperwork in several codecs – Our product documentation, accrued over many years, different enormously in format. The presence of non-textual parts, resembling tables, posed important challenges. Tables might be tough to interpret when immediately queried from PDFs or Phrase paperwork. To deal with this, we normalized and transformed these paperwork into constant codecs appropriate for the RAG system, enhancing the mannequin’s potential to retrieve and interpret data precisely. We used the FM parsing functionality of Amazon Bedrock Information Bases, which processed the uncooked doc with an LLM earlier than making a remaining chunk, verifying that information from non-textual parts was additionally accurately interpreted.
• Dealing with LLM boundaries – Consumer queries typically exceed the system’s capabilities, resembling after they request complete data, like a whole checklist of product options. As a result of our documentation is cut up into a number of chunks, the retrieval system may not return all the mandatory paperwork. To deal with this, we adjusted the system’s responses so the AI agent might present coherent and full solutions regardless of limitations within the retrieved context. We created customized paperwork containing FAQs and particular directions for these circumstances and added them to the data base. These acted as few-shot examples, serving to the mannequin produce extra correct and full responses.
• Grounding responses – By nature, an LLM completes sentences based mostly on chance, predicting the subsequent phrase or phrase by evaluating patterns from its intensive coaching information. Nevertheless, typically the output isn’t correct or factually right, a phenomenon also known as hallucination. To deal with this, we use a mixture of specialised prompts together with contextual grounding checks from Amazon Bedrock Guardrails.
• Managing one-line dialog follow-ups – Customers usually have interaction in follow-up questions which might be temporary or context-dependent, resembling “Are you able to elaborate?” or “Inform me extra.” When processed in isolation by the RAG system, these queries would possibly yield no outcomes, making it difficult for the AI agent to reply successfully. To deal with this, we applied mechanisms to take care of conversational context, enabling HxGN Alix to interpret and reply appropriately.

We examined two approaches:

• Immediate-based search reformulation – The LLM first identifies the person’s intent and generates a extra full question for the data base. Though this requires a further LLM name, it yields extremely related outcomes, protecting the ultimate immediate concise.
• Context-based retrieval with chat historical past – We despatched the final 5 messages from the chat historical past to the data base, permitting broader outcomes. This strategy offered sooner response occasions as a result of it concerned just one LLM spherical journey.

The primary methodology labored higher with giant doc units by specializing in extremely related outcomes, whereas the second strategy was more practical with a smaller, centered doc set. Each strategies have their execs and cons, and outcomes differ based mostly on the character of the paperwork.

To deal with these challenges, we developed a pipeline of steps to obtain correct responses from our digital assistant.

The next determine summarizes our RAG implementation journey.

Adjusting the appliance improvement lifecycle

For generative AI techniques, the normal software improvement lifecycle requires changes. New processes are essential to handle accuracy and system efficiency:

• Testing challenges – Not like conventional code, generative AI techniques can’t rely solely on unit assessments. Prompts can return totally different outcomes every time, making verification extra complicated.
• Efficiency variability – Responses from LLMs can differ considerably in latency, starting from 1–60 seconds relying on the person’s question, not like conventional APIs with predictable response occasions.
• High quality assurance (QA) – We needed to develop new testing and QA methodologies to be sure that Alix’s responses had been constant and dependable.
• Monitoring and optimization – Steady monitoring was applied to trace efficiency metrics and person interactions, permitting for ongoing optimization of the AI system.

Conclusion

The profitable launch of HxGN Alix demonstrates the transformative potential of generative AI in enterprise asset administration. By utilizing AWS generative AI companies and a rigorously chosen expertise stack, we optimized inside workflows and elevated person satisfaction inside safe environments. HxGN Alix exemplifies how a strategically designed AI resolution can drive effectivity, improve person expertise, and meet the distinctive safety wants of enterprise purchasers.

Our journey underscores the significance of a strategic strategy to generative AI—balancing safety, accuracy, and sustainability—whereas specializing in the proper use case and expertise stack. The success of HxGN Alix serves as a mannequin for organizations searching for to make use of AI to resolve complicated data entry challenges.

By utilizing the proper expertise stack and strategic strategy, you possibly can unlock new efficiencies, enhance person expertise, and drive enterprise success. Join with AWS to be taught extra about how AI-driven options can remodel your operations.

In regards to the Authors

Julio P. Roque is an completed Cloud and Digital Transformation Govt and an professional at utilizing expertise to maximise shareholder worth. He’s a strategic chief who drives collaboration, alignment, and cohesiveness throughout groups and organizations worldwide. He’s multilingual, with an professional command of English and Spanish, understanding of Portuguese, and cultural fluency of Japanese.

Manu Mishra is a Senior Options Architect at AWS, specializing in synthetic intelligence, information and analytics, and safety. His experience spans strategic oversight and hands-on technical management, the place he evaluations and guides the work of each inside and exterior clients. Manu collaborates with AWS clients to form technical methods that drive impactful enterprise outcomes, offering alignment between expertise and organizational targets.

Veda Raman is a Senior Specialist Options Architect for generative AI and machine studying at AWS. Veda works with clients to assist them architect environment friendly, safe, and scalable machine studying functions. Veda focuses on generative AI companies like Amazon Bedrock and Amazon SageMaker.