Drasko Profirovic is a data-platform engineer who developed Medic for Apache Spark, an AI-assisted system for diagnosing distributed computing failures and recommending practical fixes. A staff engineer at Pinterest in 2026, he focuses on making specialist operational knowledge accessible to teams that depend on complex infrastructure.
Profirovic studied at Simon Fraser University and worked on enterprise software projects for Fisheries and Oceans Canada before joining Realtor.com. There, his work encompassed AWS infrastructure, deployment tooling, GraphQL, Elasticsearch, and data pipelines. He built distributed tracing for data pipelines using Splunk, improving visibility into data moving between services. He also applied supervised machine learning to address parsing, using synthetic training examples and explicit evaluation to interpret real-estate listings that resisted handwritten rules.
At Pinterest, Profirovic applied that combination of observability, machine learning, and infrastructure experience to Medic for Apache Spark. The system investigates failed jobs, assembles evidence for likely causes, and recommends workload-specific remediation. Its original design connected diagnostic tools through Model Context Protocol and assigned investigation to a single ReAct agent, but production logs overwhelmed context windows, answers varied unpredictably, and prompt changes introduced regressions.
How Medic makes automated diagnosis reliable
- Record-and-playback agent evaluation. OpenTelemetry traces sent to Langfuse expose the agent’s execution, while recorded tool responses become reusable test fixtures. Offline evaluations assess report quality and catch regressions without depending on production data remaining available.
- Selective exception retrieval. Medic learns which exceptions also appear in successful jobs, filters those misleading signals, and ranks remaining failures by relevance and timing. Agents receive concise candidate exceptions first and request complete logs only when necessary.
- Multimodal infrastructure diagnostics. Isolated subagents convert time-series metrics into annotated visual dashboards, allowing multimodal models to identify disappearing executors, prolonged plateaus, and stalled progress without consuming the supervising agent’s context window.
- Specialized multi-agent troubleshooting. Built with LangGraph and Deep Agents, Medic separates triage, parallel investigation, root-cause selection, and runbook-based remediation. Profirovic found rigid deterministic workflows less effective than agents able to reason through changing evidence.
The architecture also supports Spark SQL optimization, with Apache Flink and Trino identified as potential future applications.
