West Hungária Bau · Construction
West Hungária Bau, a leading Hungarian general contractor, replaced manual tender pricing with a generative AI system built on Amazon Bedrock. Processing time per tender fell from 6 to 8 hours down to 1 to 2 hours, freeing over 1,000 engineering hours a year across roughly 250 tenders.
Customer: West Hungária Bau Kft. (WHB Group)
Industry: Construction, covering industrial, commercial and public infrastructure
Solution: Athene AI Pricing System
Delivery: Devertix Cloud Kft.
Key AWS service: Amazon Bedrock
West Hungária Bau Kft. is a Hungarian general contractor working across industrial, commercial and public infrastructure construction. As part of the WHB Group it competes for contracts nationwide, submitting approximately 250 tenders a year at group level.
Every tender arrived as a package of unstructured documents: customer-supplied Excel bills of quantities, technical specifications, and subcontractor offers returned in mixed Word, PDF and Excel formats. No two packages shared a structure.
Turning that package into a priced offer was manual throughout. An estimator validated the formulas in the supplied bill of quantities, built out the remaining formulas across every worksheet, read and transcribed each subcontractor offer by hand, and finally constructed the project schedule in MS Project by converting budget formulas to values and rebuilding the plan. The process consumed 6 to 8 hours of estimator time per tender and restarted from zero for every submission.
This created three compounding business risks.
Tender volume was bounded by estimator availability rather than by market opportunity. Bids arriving close to a deadline were either rushed or declined, and a declined bid forfeits the entire contract value.
Incorrect SZUM formulas, broken and circular cell references, and logical inconsistencies in the supplied workbooks were found only by manual inspection, or not at all before the tender was issued. In construction tendering, a systematic pricing error is absorbed by the contractor for the life of the project.
Estimation requires experienced engineers who understand construction methods and material substitution. Most of their tender-day hours went to reading and transcribing rather than to the judgement calls only they can make.
Without intervention, WHB faced a structural constraint on bid volume at precisely the point where its addressable market was expanding.
The Athene AI Pricing System was delivered into WHB’s own AWS account, so that commercially sensitive tender documentation and internal pricing data remain under WHB’s control and never transit systems operated by a third party. No outside party has access to WHB’s tender data.
Foundation models on AWS automate four parts of the estimation workflow:
Formula validation. SZUM formulas and cell references were previously checked by manual inspection. The system now detects syntax errors, circular and broken references, and logical inconsistencies automatically.
Formula generation. Estimators used to build out formulas by hand across every worksheet. Formulas are now constructed automatically on all worksheets, with defined exception cells deliberately left blank.
Offer processing. Subcontractor offers arriving in mixed formats were read and transcribed manually. The system now extracts entities and tabular data automatically, then maps and imports them into the priced Excel.
Schedule generation. The MS Project schedule used to be built by hand. It is now generated automatically from recognised budget line items, following the schemas of earlier project files.
Amazon Bedrock provides the foundation-model capability the automation depends on. Without it there is no formula reasoning, no mixed-format document comprehension and no budget line-item recognition. The four functions above are not separable from it.
Bedrock was selected over self-hosting a model on Amazon SageMaker or on EC2 GPU instances. Self-hosting was evaluated and rejected on two grounds. Tender arrival is bursty, with quiet weeks punctuated by deadline clusters, so a provisioned inference endpoint would have been idle for most of its billed lifetime. And managed model serving removed an MLOps burden from a customer whose IT function is not staffed for it. Bedrock is consumed in an EU region, keeping inference within European data-handling commitments.
Amazon Bedrock is the foundation-model engine behind formula reasoning, document comprehension and budget line-item recognition, and it is invoked under WHB’s own AWS subscription.
Tender documents and mixed-format subcontractor offers entering the extraction pipeline are stored in Amazon S3, which also hosts the web frontend. Amazon CloudFront delivers that frontend and is the only component reachable from the public internet, with AWS Certificate Manager managing the TLS certificate for the public endpoint.
The application services run as containers on Amazon EKS, with Karpenter provisioning nodes across multiple Availability Zones. Application data and offer history, which allows fallback to previous subcontractor offers, are held in Amazon Aurora Serverless v2, chosen because it scales with tender activity rather than being sized for peak load.
Amazon OpenSearch Service indexes and retrieves content across processed documents and offer history. Relational full-text search was tested first and could not cope with the vocabulary variation between issuing authorities. Amazon CloudWatch provides operational monitoring, logging and threshold-based alerting.
The solution runs across multiple Availability Zones within a VPC. EKS workloads and Aurora 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. Scaling is automatic throughout: Karpenter for EKS compute, and native demand-based scaling for CloudFront, S3, Lambda and Bedrock. This matters specifically because tender deadlines cluster: the system must absorb a week’s volume in a day without capacity planning by the customer.
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 in S3, Aurora and EFS 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 pipelines into Argo CD. No production change is made through the AWS Management Console. Delivery ran as an agile programme with named sprints, sprint planning, daily stand-ups and demo sessions held with the customer, so functionality was reviewed by the business as it was built. Acceptance was tracked as explicit work items rather than assumed. Once released, functionality moves into a dedicated service management space for the operational phase.
During the discovery phase, the customer put manual processing time at 6 to 8 hours per tender, and the project goal was to save 4 hours on each one. The post-launch review with the customer confirmed an actual processing time of 1 to 2 hours per tender, a reduction of approximately 75%, meeting and exceeding the target.
WHB submits approximately 250 tenders a year at group level, a figure provided by the customer during discovery. At the baseline processing time, that volume consumed approximately 1,500 to 2,000 hours of manual estimator effort annually. The target, derived from the per-tender goal, was to release approximately 1,000 hours a year, a reduction of 50 to 65%. Applying the measured per-tender reduction to annual volume, the solution now releases approximately 1,000 to 1,500 engineering hours a year.
Those hours return to estimation judgement, and to bidding on opportunities previously declined for lack of time.
Alongside processing time, the solution changed the reliability of the underlying calculations. Automated formula validation catches incorrect SZUM formulas, broken and circular cell references and logical inconsistencies. These are defects that were previously found only by manual inspection, or not at all before the tender was issued. This outcome has no baseline figure, because the defects it prevents were not being counted before.
The system runs in WHB’s own AWS environment under an ongoing paid support arrangement. After go-live, WHB commissioned a further phase extending the generative AI data-extraction capability, which is the customer’s own validation of the business case.
A cost model was produced for WHB before implementation, so that the ongoing run cost was a known input to the investment decision rather than a surprise after go-live.
The model was built from WHB’s actual operational profile: annual tender volume, average document count and page count per tender package, the resulting Bedrock token consumption for extraction and reasoning, storage growth for retained tender archives, and the compute baseline required to keep the application available between tender peaks. Because the workload is bursty rather than steady-state, the model was built around consumption during active tendering periods rather than a flat monthly average. That is what made the serverless and consumption-priced service choices defensible on cost as well as on architecture.
The estimate was presented to WHB alongside the architecture proposal, broken down by service so that the customer could see which components drive cost and how each scales with tender volume. Actual running cost has tracked the model.
Against the run cost, the value of released estimator capacity was modelled: the hours returned, valued at WHB’s loaded engineering cost, together with the option value of tenders that could now be bid within deadline rather than declined. The efficiency saving alone covers the platform’s operating cost by a wide margin; the bid-capacity effect is the larger but less directly measurable component.
The initial extraction approach was built against a representative sample of tender packages. Live traffic contained format variations such as scanned inserts, non-standard bill-of-quantity layouts and issuer-specific abbreviations that the sample did not cover, and extraction accuracy on those outliers fell below target in early production. A document-variability assessment across a materially larger sample is now run during the scoping phase, and the prompt-refinement effort is sized accordingly rather than treated as a fixed task.
Extraction accuracy improved measurably through iterative refinement after go-live, driven by estimator feedback on mismatched items. Prompts are now versioned in Git alongside application code and prompt updates are treated as a reviewed, released change, with a structured feedback path from end users established from day one rather than relying on ad-hoc reports.
The estimator’s approval step was initially framed as a confidence-building measure to be relaxed over time. It has proven to be the correct permanent design: it keeps accountability for a priced offer with the person who owns it commercially, and it generates the feedback that drives accuracy improvement. Human-in-the-loop review is now designed into generative AI solutions as a standing control.
Source control and CI pipelines for both application and infrastructure code run on GitHub and GitHub Actions, and Argo CD handles GitOps continuous delivery into Amazon EKS. 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.
Delivery was managed in Atlassian Jira Cloud, covering sprint tracking and user acceptance testing, while Atlassian Jira Service Management handles operational support and incident handling for the live service. Project and solution documentation is kept in Atlassian Confluence, and Microsoft Project is the target output format for the generated construction schedules.
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.
Athene AI, the pricing system delivered to WHB, is Devertix’s own product. More at https://www.atheneai.hu/en-us