Stratasys cuts AI query costs 9x and delivers product answers 6x faster

Stratasys supports a broad industrial manufacturing portfolio that includes 3D printers, materials, and software. Its product knowledge spans manuals, technical specifications, knowledge articles, training videos, and spare parts catalogs. Customer support engineers and sales teams need to access this information quickly, including during live conversations with customers across the automotive and manufacturing industries.

However, the information was distributed across multiple repositories, available in different formats, and governed by role-based permissions. This made it difficult for traditional search tools to consistently surface accurate and relevant information from thousands of digital assets.

During an early PoC, Stratasys experimented with an AI knowledge assistant on Amazon Bedrock using a frontier foundation model. The assistant delivered strong answer quality, but the selected model and pipeline design resulted in projected annual costs of more than $600,000 for a single component at full production scale. Response times also exceeded the requirements of an interactive support tool, where engineers often need answers within seconds. These constraints made it difficult to expand the solution to partners and end customers.

“We needed a way to deliver accurate answers at a lower cost while making the solution fast enough for real-time use across our global teams,” says Amir Kleiner, chief operating officer at Stratasys.

Stratasys worked with AWS Partner Karini AI to build an AI knowledge assistant with role-based access on the Karini AI platform. The solution uses Amazon Nova Multimodal Embeddings to index product manuals, technical documents, training videos, and product images. Amazon OpenSearch Service supports vector and keyword retrieval, while Amazon Neptune maps product relationships to help users find information such as compatible spare parts.

For the question-answering experience, Karini AI fine-tuned Amazon Nova 2 Lite using Amazon SageMaker AI serverless training. The team developed a training dataset from real-world questions and answers captured through Karini AI’s observability layer and validated through user feedback.

Karini AI then used Amazon Nova Premier as a teacher model to generate consistent training samples through Amazon Bedrock batch inference, reducing data preparation costs by 50 percent. When the team identified question categories that needed greater coverage, it expanded the dataset and completed a second fine-tuning iteration. The fine-tuned Amazon Nova 2 Lite model is now deployed on Amazon Bedrock for on-demand inference.

The same observability layer captures new queries, responses, and user feedback after deployment. Stratasys and Karini AI use these insights to identify gaps, refine the training data, and improve the assistant over time. Each deployment cycle builds on real-world usage, helping improve answer quality while reducing the manual effort required to prepare training data.

By fine-tuning Amazon Nova 2 Lite using Amazon SageMaker AI, Karini AI helped Stratasys move from PoC to production deployment. The AI knowledge assistant now supports more than 2,000 users worldwide.

Compared with the frontier foundation model used in the initial PoC, the fine-tuned model reduced cost per query by approximately nine times. End-to-end response times were also approximately six times faster, helping engineers access information during live customer conversations. These improvements supported a broader rollout across customer support, sales, marketing, and partner teams.

On Stratasys’s evaluation dataset, answer quality approached the baseline established by the original frontier model. The team assessed correctness, completeness, and answer structure using an LLM-as-a-judge methodology. The assistant now supports 12 languages across Stratasys’s global operations.

The project also established a repeatable fine-tuning approach that Stratasys is applying to other parts of the pipeline, including knowledge graph query generation and metadata filtering. These efforts create further opportunities to improve answer quality and reduce operating costs.

Nitin Wagh, CEO and co-founder of Karini AI, says, “By fine-tuning Amazon Nova 2 Lite using Amazon SageMaker AI for Stratasys’s AI knowledge assistant, we delivered the answer quality Stratasys needed at a cost that made production deployment viable. The combination of Amazon Nova’s performance and the Karini AI platform’s observability provides a continuous improvement loop that uses real-world feedback to refine the solution over time.”