The positions behind every architecture decision.
How we think about agentic AI, knowledge graphs, and operational engineering — the convictions we bring to every engagement, earned by building, deploying, and running these systems.
Agents that ship are designed as operations, not demos.
Most agentic systems collapse the moment they leave the demo. Production autonomy requires explicit design contracts — bounded authority, observable decisions, evaluated behavior, and a human escalation path. We architect all four in from day one.
Discuss this with the architectarrow_forwardConvictions we'll defend on a whiteboard.
Ontologies are an operating contract, not a data project.
Ontology design is a senior architecture activity. Treated as a one-off data exercise, it dies at rollout; treated as a contract, it compounds.
Instrument decisions, not just calls.
Traditional APM tells you what happened. Agentic observability has to tell you why the agent decided to do it — or you can't operate it.
Orchestration frameworks are a thin, replaceable boundary.
Teams over-index on the framework du jour. We keep orchestration thin so the architecture outlives whatever library is fashionable this quarter.
No evaluation harness, no production.
If regressions can't be caught before they ship, the system isn't done. Continuous evaluation is part of the build, not a follow-up phase.
Compliance is an architectural property.
'We'll add compliance later' is a load-bearing decision in disguise. Security and auditability get baked in from week one or paid for forever.
When vectors aren't enough, go graph-native.
There's a class of enterprise questions — multi-hop, relational, governed — where graph retrieval substantially outperforms pure vector search.
Want one of these expanded for your context?
Long-form field notes are being drafted. If a position above maps to a problem you're facing, skip the wait — bring it to a conversation.