Mediagnost · Precision Manufacturing
Mediagnost, a Hungarian manufacturer of mechanical and diagnostic structures, replaced the manual annotation and transcription of technical drawings with a generative AI application running on Amazon Bedrock inside its own AWS account. Engineers now validate output instead of transcribing it, and confidential drawings never leave systems under the customer’s control.
Customer: MEDIAGNOST Mechanikai és Diagnosztikai Szerkezeteket Gyártó Kft.
Industry: Manufacturing, mechanical and diagnostic structures
Solution: AtheneAI platform, drawing-ballooning application and web search extension
Delivery: Devertix Cloud Kft. (direct contract)
Key AWS service: Amazon Bedrock with Anthropic Claude models
MEDIAGNOST Mechanikai és Diagnosztikai Szerkezeteket Gyártó Kft. is a Hungarian manufacturer of mechanical and diagnostic structures, based in Budapest. Its engineering and quality functions work from technical drawings and confidential internal documentation, producing precision components whose every dimension, note and symbol must be recorded in customer-mandated quality reporting formats.
Technical drawing documentation was a per-item manual process. For each drawing, an engineer had to annotate every dimension, note, drawing symbol such as surface roughness or welding, and every material or surface-finish entry in the text field. Each item required a numbered marker and a hand-drawn pointer. The engineer then read each annotated value back off the drawing and transcribed it into the customer’s SRPQP Excel template, one row per item, keyed by drawing number and balloon number.
The defining property of this process is that effort scales multiplicatively: features on each drawing multiplied by the number of drawings processed. It is not a fixed overhead that dilutes as volume rises. It grows in direct proportion to production, so engineering capacity becomes a constraint on the business at precisely the moment the business wins more work.
A misread tolerance value does not stay inside the engineering function. It propagates into quality documentation that Mediagnost’s own customers rely on, and it is discovered downstream rather than at source.
Knowledge held across internal documentation could only be reached by manual search. Making it queryable through an externally hosted AI service was not an option, because it would have meant placing confidential engineering material outside the company’s control.
The engagement did not begin with a technology build. Following an initial proof of concept, a structured generative AI adoption process was delivered as dedicated advisory, designed to find where AI would create value in Mediagnost’s own operations rather than to justify a predetermined solution.
The process ran in four phases.
Strategic kick-off and AI-readiness assessment. A workshop with nominated key users and decision-makers presented the platform’s capabilities in business rather than technical terms and identified operational pain points such as repetitive tasks, knowledge-sharing gaps and slow processes. It produced a shared view of where AI could help most.
Use case ideation and prioritisation. Brainstorming workshops explored the areas identified, and every candidate use case was evaluated against an Impact and Effort framework. The result was a prioritised list of Quick Wins and higher-ROI use cases.
Quick Win pilot definition. A technical deep-dive established the data and integrations required, while a business deep-dive defined the success measures and how productivity and user satisfaction would be measured. Both were captured in a Pilot Charter for a three-month pilot.
Closing and roadmap. The Pilot Charter was presented to management, together with the training and development steps to follow, securing management commitment and an agreed roadmap.
The drawing-ballooning application is the Quick Win that came out of this process. It was selected because it addressed a repetitive, high-volume engineering task, not because it was the most technically interesting thing to build.
Generative AI was introduced progressively rather than as a single cutover. Each stage was commissioned only after the previous one had proven itself in use.
The engagement opened with an AtheneAI proof of concept that established feasibility and the technical approach before any platform commitment was made. A Starter Package followed, deploying the AtheneAI platform into Mediagnost’s own private AWS environment together with dedicated AI advisory.
On that foundation Devertix built the drawing-ballooning application, which annotates technical drawings and extracts the annotated data into the SRPQP Excel reporting format. A web search extension came next, enabling the platform to answer from current web sources rather than only from the foundation model’s trained knowledge. Continued operation and development are now supported through a standing AI and DevOps advisory arrangement.
The AtheneAI platform runs inside Mediagnost’s own AWS environment. This is the architectural premise of the engagement rather than an implementation detail. The customer’s requirement was that corporate data, documents and source code never leave systems under its own control, which ruled out externally hosted AI services and made in-account foundation-model access the enabling capability.
Knowledge access. Engineers previously had to search manually through confidential internal documentation. They now ask questions in natural language against a project-specific knowledge base held inside Mediagnost’s own AWS environment.
Drawing annotation. Numbering and pointer drawing used to be done by hand for every dimension, note and symbol. The application now annotates drawings automatically, with configurable pointer direction, adjustable size categories and recognition of diameter-type dimensions.
Data extraction. Each annotated item was once read off the drawing and typed into Excel. The application now extracts the data in structured form into the SRPQP template, one row per item, and exports the result as SRPQP_[drawing number].xlsx.
Current information. The model was previously limited to its trained knowledge, and any external research was done manually. With web search integrated into the application, answers can now draw on current sources.
The ballooning specification was written from how Mediagnost actually works. It sets out precisely which drawing content must be annotated and which must not, with the parts list and item numbers explicitly excluded. It records the customer’s existing annotation convention, the customer’s own SRPQP Excel template and its column semantics, and the required output file-naming convention. Real customer drawings and correctly completed Excel outputs were supplied as reference material.
The same specification records the questions put to the customer during discovery: the current manual workflow and its volumes, the intended users, who validates AI output and on what interface, the consequences of a misread tolerance value, and how ambiguous data should be flagged. Human-in-the-loop validation was treated as a design requirement from the outset, not as a transitional safeguard.
Amazon Bedrock is the foundation-model engine of the solution. It provides Anthropic Claude models for natural-language processing, document comprehension and drawing interpretation, and users can choose the model per conversation, with Claude Sonnet offered as the standard option.
Drawings and source documents entering the processing pipeline are stored in Amazon S3, which also hosts the web frontend. The frontend is delivered through Amazon CloudFront, the only component reachable from the public internet, and AWS Certificate Manager manages the TLS certificate for that public endpoint.
The application services run as containers on Amazon EKS, where Karpenter provisions nodes across multiple Availability Zones. Application data is held in Amazon Aurora Serverless v2, deployed multi-AZ for fault tolerance and scalable capacity, while Amazon OpenSearch Service, also spread across multiple Availability Zones, handles indexing and retrieval across the project-specific knowledge bases. Amazon CloudWatch provides operational monitoring, logging and threshold-based alerting.
The solution runs across multiple Availability Zones within a VPC. EKS workloads, Aurora and OpenSearch are deployed in private subnets spanning several AZs, with a Network Load Balancer routing only to healthy targets. S3, CloudFront, Route 53, Lambda, Bedrock and CloudWatch provide availability as managed services. EKS compute scales through Karpenter, and CloudFront, S3, Lambda and Bedrock scale natively with demand.
Only the CloudFront distribution is reachable from the public internet. All other components sit behind it on internal network paths, each with its own security group permitting only required ports. SSL terminates at the Network Load Balancer. All data at rest is encrypted with AWS Key Management Service.
Infrastructure is defined as code using OpenTofu with Terragrunt for environment-specific configuration. Application releases flow through GitHub Actions and GitLab CI/CD pipelines into Argo CD. No production change is made through the AWS Management Console.
The generative AI application performs the annotation and the structured extraction, and produces the populated SRPQP Excel file directly. The engineer’s role changes from producing every row by hand to checking the rows that were produced. Because effort previously scaled with features per drawing times drawings processed, removing the per-item step breaks the link between production volume and engineering workload.
Success measurement was defined during the advisory stage rather than added afterwards. The business deep-dive explicitly required the engagement to answer what a successful pilot looks like and what the measures are, with productivity and user-satisfaction measurement customised for the selected use case and recorded in a Pilot Charter covering a three-month pilot horizon.
The clearest evidence of value is what Mediagnost did next. The company moved from a proof of concept to a platform deployment, commissioned two separate production feature builds, and placed the relationship on a standing monthly advisory arrangement.
One commission is particularly indicative. It was placed not to build the ballooning application but to enhance one already in operational use, and its content is a list of refinements arising from that use: more precise diameter recognition, adjustable pointer direction, editable size categories, bulk deletion, and constraints preventing annotations from overrunning the page. Requests of that specificity come from users working with the tool daily.
The application remains in operational use, supported under a standing advisory arrangement.
Structuring the engagement so that each stage had to earn the next meant the customer never committed budget ahead of demonstrated value, and Devertix never built ahead of validated requirements. For an organisation adopting generative AI for the first time, this proved more effective than a single larger programme, and it is now the default shape for first AI engagements.
The enhancement list, covering diameter recognition precision, pointer direction, size categories, bulk deletion and page-overrun constraints, could not have been written before the application was in daily use. A refinement phase after go-live is now planned into engagements of this type from the start, rather than treated as rework.
Who validates AI output, on what interface, and how ambiguous data is flagged were discovery questions rather than build-phase decisions. Where a misread value reaches a customer’s quality documentation, the validation step is part of the product, not a temporary control to be relaxed once confidence grows.
Source control and CI pipelines run on GitHub and GitHub Actions, with GitLab CI/CD providing additional build and release pipelines. Argo CD handles GitOps continuous delivery into Amazon EKS, and infrastructure is defined as code with OpenTofu, using Terragrunt for environment-specific configuration.
Inside the cluster, Karpenter provisions and autoscales nodes, Traefik handles ingress routing, and cert-manager manages the TLS certificate lifecycle for services behind the Network Load Balancer. Microsoft Excel is the target output format, matching the customer’s SRPQP quality reporting template.
Devertix Cloud Kft. is a Hungarian IT company and AWS Advanced Tier Services Partner, specialising in AWS cloud services, DevOps, custom application development and generative AI solutions.
AtheneAI is developed by Devertix Cloud Kft. together with its partners Alvicom Kft. and R-Szoft Kft. More at https://www.atheneai.hu/en-us