Work

Ongoing platform work inside a Czech enterprise software company: multi-tenant services, integration with customer systems, data pipelines and the reliability work that keeps them running. The hardest part was never the code — it was that enterprise customers change requirements and infrastructure simultaneously, and the system has to absorb both.

• Multi-tenant service architecture serving several distinct customer deployments
• Integration layer for CRM, ERP and billing systems with retries and reconciliation
• Deployment pipeline and environments matching the customer's security requirements

A content platform with a public API, rebuilt around typed contracts, caching and predictable latency. The interesting constraint was the opposite of AI work: nothing could change that visitors and integrations could see. Every field, endpoint and error format was versioned deliberately.

• Typed request and response contracts across the whole public surface
• Caching layer that removed the database from the read path
• Editorial workflow improvements that cut publishing time measurably

Battery diagnostics for fleets and workshops: device fleet management, telemetry ingestion, threshold alerting and reporting. Volumes are spiky and unpredictable — which made backpressure, idempotent writes and honest alerting the actual engineering problem.

• Device fleet management with provisioning and firmware state
• Telemetry ingestion that survives intermittent connectivity
• Diagnostic thresholds and reporting for workshop workflows

Client work in regulated environments: hybrid retrieval over internal documentation, reranking, citations, permission filtering, and an evaluation harness that runs in CI. The measure of success was never demo quality — it was whether the audit team accepted the answers.

• Hybrid BM25 + vector retrieval with cross-encoder reranking
• Time-sliced access filters applied at query time, never in the prompt
• Golden datasets from real traffic, scoring every release

Website assistants grounded in client documentation, answering in the visitor's language with citations and human handoff. Instrumentation ties conversations to support-ticket volume, so the client can see deflection instead of assuming it.

• One knowledge base, three languages, per-language quality reporting
• Explicit refusal behaviour with structured escalation to a human
• Privacy-first analytics: no third-party trackers, no cookies for chat logging