Our research.
Peer-reviewed publications by the LabSoft team — AWS architecture, AI agents, serverless computing, and software engineering.
Why we publish
Production AI is a young field. Most of what works in deployed systems is folklore — passed between teams, never written down. We publish what we learn: what scales, what doesn't, where the cost curves bend, when to use Docker vs native runtimes on Lambda. The papers below are the same engineering decisions we apply on client engagements.
Comparison of AWS Architectures for Scalable and Cost-Efficient Retrieval-Augmented Generation
Dimitrije Stojanović, Luka Vidaković, Bogdan Pavković, Nenad Četić, Momčilo Krunić
Large Language Models (LLMs) require up-to-date and domain-specific knowledge to generate accurate responses. We present a serverless RAG architecture on AWS that leverages Lambda, S3, DynamoDB, and API Gateway to automate scaling and implement a pay-per-use model. Our evaluation demonstrates savings of up to 87% for loads of 10,000 requests per hour compared to EC2 instances.
Comparative Analysis of Docker and Python Runtimes for AWS Lambda in RAG-Based AI Solutions
Luka Vidaković, Dimitrije Stojanović, Bogdan Pavković, Nenad Četić, Momčilo Krunić
We provide a thorough comparison between Docker-based and Python-native runtimes on AWS Lambda, examining cold start latency, warm execution time, and build time. Our empirical findings show that the Python-native runtime achieves up to an 84% reduction in execution times for RAG-based AI solutions.
Unit Test Generation Multi-Agent AI System for Enhancing Software Documentation and Code Coverage
Dimitrije Stojanović, Bogdan Pavković, Nenad Četić, Momčilo Krunić, Luka Vidaković
We explore the utilization of AI agents for generating and executing unit tests, enhancing the "Mostly Basic Python Problems" dataset. We employ behavior-driven development within a three-agent system to generate user stories and unit tests. Empirical results indicate improvements in branch coverage.
Documentation as Code in Automotive System/Software Engineering
Momčilo V. Krunić
Documentation as Code (DaC) applies software development principles to technical documentation. This paper discusses DaC advantages in automotive system/software engineering including improved accuracy, traceability, and maintainability. Research conducted with 150+ engineers actively contributing to DaC for over a year.