From 14 Days to Minutes — Generative AI Transforms HR & Accounting

At Cosmos Aluminium, two departments were losing significant time to slow, manual information retrieval. The HR team routinely managed more than 1,000 résumés, each formatted differently and using inconsistent terminology. Screening candidates for an open position meant manually reviewing documents one by one — a process that took an average of 14 days.

Meanwhile, the accounting team relied on 170 SAP ERP manuals to guide day-to-day processes. Employees needing guidance had to manually search those documents, ask colleagues, or submit help desk tickets — often waiting up to 30 minutes for an answer. This slowed decision-making and introduced inconsistencies across workflows.

"The main scope of this project is to search quickly. The process was manual and slow, relying heavily on individual experience. This often caused delays and inconsistent results." — Petros Chondronasios, IT Manager, Cosmos Aluminium.

Building on an existing relationship, Cosmos Aluminium selected LCM Go Cloud for their expertise in delivering generative AI on AWS. The team built two specialised AI agents on Amazon Bedrock — each targeting one department's problem precisely. The CV Intelligence Application evaluates résumés against job requirements using Claude Sonnet and Haiku models, identifying similar skills even when described in different terminology. The Accounting Manual Application searches all 170 SAP ERP manuals to answer natural-language questions instantly, in both Greek and English.

The architecture processes documents asynchronously across 14 Lambda functions and 6 queues, stores raw and processed files in 5 Amazon S3 buckets, and uses Amazon DynamoDB for metadata and session history. Amazon Cognito ensures users can access only information relevant to their role. Security controls include encryption at rest and in transit, antivirus scanning of all uploads, and data residency strictly within the EU.

The entire project — from kickoff to production — was completed in just 16 weeks. From day one, 25 users adopted both applications with no formal training beyond a brief orientation. The solution exceeded its 95% accuracy baseline, with real-world performance closer to 99% accuracy across both systems.