About
I build the evidence layer around AI systems—evaluation harnesses, deterministic safety gates, fault injection, and trace reconstruction.
“Build systems that fail loudly, preserve the evidence, and block unsafe action before a confident mistake reaches a system of record.”
I came to AI systems through manufacturing quality engineering. In industrial manufacturing, correctness is not a stylistic preference: a bad decision instantly impacts physical inventory, factory lines, or general ledger balances.
When teams deploy Large Language Models into financial or enterprise workflows, they often assume that because the output parses as JSON, the output is true. But schema compliance only guarantees syntax—it says nothing about whether the cited document exists, whether the tax calculation balanced, or whether an adversarial prompt manipulated the model's instructions.
I focus on building the deterministic shell around probabilistic models: contracts that enforce preconditions and postconditions, byte-reproducible test corpora, and safety gates that isolate failures before they reach the real world.
Track Record
Engineering Principles
Independent truth
A detector must never grade itself. I evaluate failures against deterministic or human-owned ground truth, and keep schema validity completely separate from semantic correctness.
Reproducible claims
Seeds, configuration, test datasets, traces, and statistical methods must travel with the result. If another engineer cannot reproduce it from scratch, it is not evidence.
Guarded production writes
AI output never earns write authority by being fluent. Writes into ERP, financial, and inventory ledgers pass deterministic verification gates or escalate to human review.
Operational ownership
I care about the system after cutover: observability, incident reconstruction, latency budgets, rollback behavior, and the total cost of recovery.
Flagship System
Reliability workbench for document-grounded AI. Deterministic fault injection, independent evaluation, trace reconstruction, and replay with 302 automated tests and zero vendor lock-in.
Collaboration
I am looking for high-ownership AI reliability, evaluation infrastructure, and agent systems engineering roles. Reach me directly on X or by email.