👋 I founded Plexito. Plexito builds custom software, AI agents and workflow automations for businesses, and runs them afterwards: hosting, maintenance and support in one monthly price.
By day I lead Technical Account Management for UiPath in DACH, where enterprises put AI agents into production. Before that, I brought retailer APIs at Instacart from 0-1 and scaled Delivery Hero logistics 1-100.
Building: spantruth. Does your LLM telemetry tell the truth? It replays provider responses through instrumentation libraries and checks every token count, model and error they report against the wire. Nightly report.
Latest essay: The agent that approved itself. An agent wrote a human approval into its own output, so the gate has to live in the workflow, not in the prompt.
Recent fixes upstream:
- mlflow#26394: keep tool and agent names on GenAI
execute_toolandinvoke_agentspans - haystack#13047: parse empty tool call arguments in the OpenAI generators
- promptfoo#11381: keep
$sequences literal when red-teaming with datamarking - docling#4295: fall back to the HTML body when an email's plain part is blank
- unstructured#4501: reject a negative chunk overlap instead of corrupting text