Writing

Notes from the overlap of systems engineering and AI execution.

These are the themes I keep returning to: durable agent behavior, boring production rollouts, and backend systems that stay understandable under pressure.

Continue the conversation
Working note

Why most internal AI support bots fail after the demo

The problem is usually not the model. It is weak retrieval, shallow tool contracts, and no disciplined escalation boundary when the bot leaves the happy path.

  • Answer from evidence first, not from vibes
  • Treat tools as product surfaces with contracts, retries, and observability
  • Escalation is a design feature, not an admission of failure
Working note

What migrating backend services to Go actually buys you

The win is rarely just throughput. The bigger gain is simpler operational behavior, clearer ownership, and a codebase the team can reason about during incidents.

  • Reduce moving parts before you benchmark hero numbers
  • Migration succeeds when API edges and rollout plans are boring
  • Operational confidence matters as much as language preference
Working note

Building agent platforms without lying to yourself

If the plan is fuzzy, the execution graph should not be. Typed interfaces, durable state, and explicit human checkpoints are what make agent systems trustworthy.

  • Use typed execution units where possible
  • Persist enough state to recover without duplicated side effects
  • Keep humans in the loop where cost or risk is asymmetric