AI AgentsVolvo Cars
LLM-as-Judge Evaluation Framework
An evaluation framework for a production AI agent. Each test case defines natural-language criteria, and a configurable judge model scores both the agent's final output and its full session trace. Replaced manual side-by-side model comparison with repeatable batch runs across ~100 test cases, and is now the team's standard for model selection and feature validation.
LLM-as-JudgeAgent EvaluationPrompt EngineeringLangChain4j
Data StreamingVolvo Cars
Real-Time Agent Session Pipeline
A Kafka pipeline that streams every agent interaction (user and AI messages, tool calls, and tool responses) to analytics as it happens, captured by decorating LangChain4j's ChatMemory without touching agent logic. At-least-once delivery with deduplication keys, retries, and a dead-letter topic replaced a weekly manual Redis-to-Snowflake export that risked data loss.
KafkaLangChain4jJavaVirtual ThreadsEvent Schema Design
AI AgentsVolvo Cars
Per-Agent Failure Handling
An exception-handling layer for a multi-agent chatbot where errors used to reach users as raw 5xx failures. It attributes each error to its originating agent while preserving the session trace, applies per-agent rules for retries, fallbacks, and messages, and returns structured responses so users always get a meaningful answer.
Multi-Agent SystemsAgent ReliabilitySession Tracing
InfrastructureVolvo Cars
Zero-Downtime Keycloak & Kafka Migrations
Moved identity management from a legacy Keycloak to a self-managed deployment on Kubernetes, with dual-server support so the application could authenticate against both during cutover; authentication response times dropped by about two-thirds. Also moved Kafka services from AWS to Confluent Cloud on Azure with no downtime and no data loss.
KeycloakKotlinKubernetesKustomizeKafkaConfluent CloudAzure
Backend SystemValitive AB
Company Data Unification Pipeline
Data-source integrations for an event-driven pipeline that unifies company and business data from disparate sources, transforming ~12M raw records into structured company entities. Designed the Elasticsearch index and query layer that powers the customer-facing search and lookup product.
KafkaAkka StreamsElasticsearchProtocol BuffersKubernetesAWS
Core BankingTosan
Bank API Integration Platform
Led a team of 6 engineers through an end-to-end initiative to build a platform that lets third-party providers connect to core bank services, and architected the banking service infrastructure behind it. Adopted by 4+ financial institutions.
JavaRESTMicroservicesTeam Leadership