I build applied AI systems that have to survive messy documents, misleading tests, and real business rules.
An invoice workflow that uses Gemini to extract structured data, applies review rules chosen by the account controller, and sends approved bills to QuickBooks Online or Xero. I completed Google's CASA Tier 2 security assessment and brought the product to SOC 2 readiness.
A mutation-testing tool for agent evals. It plants controlled failures in passing traces, reruns the same evaluators, and shows teams which broken tool calls, repeated side effects, hidden errors, and damaged responses their tests miss.
View Mendmark · Read the field note
A Python package that turns trained, one-output LightGBM models into dependency-free Python, C++17, or JavaScript. The test suite executes or compiles every target and compares its raw scores with LightGBM at 1e-12 tolerances.
View lgbm-to-code · Read the case study · View on PyPI
I start with the failure that would make a clean result misleading. Then I build the data, tests, workflow, and recovery path around it.