Hamza Tahir is a co-founder of ZenML, the company behind the open-source machine-learning infrastructure framework ZenML and the agent-evaluation project Kitaru. He builds tools that make production AI systems reproducible, inspectable, and easier to improve.
From machine-learning pipelines to AI agents
Tahir started in software development before moving into machine-learning engineering and co-founding maiot, where work on production applications exposed the operational difficulty of assembling reliable ML infrastructure. His earlier independent projects included Scrilys, a browser-based writing application.
In 2020, he co-authored a TensorFlow Blog article on distributed principal-component analysis, describing how TensorFlow Extended, TensorFlow Transform, and Apache Beam could apply the same data transformation during training and serving. That consistency helped prevent production models from encountering differently prepared inputs.
Tahir and Adam Probst subsequently built ZenML into an open-source MLOps framework for reproducible pipelines spanning different orchestrators, cloud providers, and infrastructure components. As agent-based applications emerged, Tahir applied similar operational principles to Kitaru: preserve execution context, replay realistic scenarios, and evaluate proposed changes before deployment.
His distinctive approach to agent reliability
Save executable state, not merely telemetry. Conventional traces document prompts, tool calls, and outputs but can omit code, runtime environments, intermediate artifacts, and application state. Tahir’s approach to durable agent execution checkpoints those missing elements so engineers can recover and inspect the conditions behind an agent’s decisions.
Replay meaningful alternatives. With checkpointed execution replay, teams can resume a saved run, substitute a model, modify a policy, or mock a tool without repeating already-preserved steps. In a customer-support agent demonstration, Tahir compares branches using GPT-5 nano and an altered policy, exposing changes in decisions, artifacts, cost, and execution time.
Turn production failures into evaluations. Tahir’s Kitaru evaluation workflow converts production traces into replayable scenarios, expert-reviewed cohorts, calibrated evaluators, and repeatable regression tests. This production-grounded agent evaluation helps teams assess prompt, model, and code changes against cases drawn from actual operations.
Judge model changes across representative cohorts. A cheaper model can look successful in an isolated replay while resolving fewer customer problems overall. Tahir advocates cohort-based model evaluation, grouping expensive, risky, or slow executions and comparing interventions across many cases; JSON reports and MCP-based analysis help determine whether apparent savings justify the operational tradeoffs.
