Vanč Levstik is a senior engineering manager at Phaidra, building production-ready AI systems for data centers and other complex industrial infrastructure. His work addresses a fundamental problem in industrial AI: language models cannot reliably identify equipment, retrieve telemetry or enumerate assets when facilities contain hundreds of thousands of nearly indistinguishable sensors and machines.
Earlier in his career, Levstik contributed to Apache Libcloud, implementing a SoftLayer DNS driver and improving integrations with Amazon Route 53 and HostVirtual, as recorded in the project’s official changelog. He subsequently helped build the team at Noala, a technology company focused on speech and language therapy.
At Phaidra, Levstik concentrates on production readiness for industrial AI. Working alongside AI research engineer Raahul Singh, he tests whether infrastructure-querying systems remain accurate under real customer data, realistic operational loads and increasing equipment counts. Their AI Engineer presentation describes evaluations spanning installations with 64 to 460,000 GPUs; across 66 cases involving six production systems, their redesigned approach recorded no failures and maintained approximately 9,000 tokens per query.
- Semantic blindness at industrial scale: Similar equipment identifiers, sprawling inventories and repetitive enumeration can defeat vector search, saturate context windows and produce nonexistent or missing assets.
- Bounded orchestration and deterministic retrieval: Models interpret ambiguous requests and generate search plans; indexed infrastructure hierarchies, filters and set operations resolve the actual equipment reproducibly.
- AI-native software matures toward deterministic code: Prompt-driven prototypes reveal what to build, but exact matching, counting and structured retrieval move into conventional software as reliability requirements sharpen. Language models retain responsibility for interpretation, judgment and readable answers.
