Insights

    Practical AI thinking for people shipping real systems.

    Short, opinionated field notes on RAG, automation, healthcare AI, product architecture and the gap between a convincing demo and a reliable workflow.

    NOTE 01
    RAGEnterprise AIEvaluation

    What makes an enterprise RAG copilot trustworthy?

    A useful enterprise copilot is not a generic chat interface connected to a folder. Trust comes from controlling how knowledge enters the system, how retrieval behaves and what the interface does when evidence is weak or contradictory.

    Read full field note Related service

    Retrieval before generation

    If the right evidence is not retrieved, a better prompt rarely fixes the underlying problem.

    Citations must support the claim

    A source link is useful only when it actually supports the sentence the user is relying on.

    Conflicts need a workflow

    Policies change. Product pages disagree. The system should surface conflicts rather than silently choose.

    Evaluation is continuous

    Golden questions, adversarial cases and regression checks should evolve with the knowledge base.

    NOTE 02
    AutomationAgentsOperations

    Automation vs AI agents: where should a business start?

    The useful question is not whether a workflow should use an agent. It is which steps are deterministic, which require judgment, which require language and which still need a person to approve the outcome.

    Read full field note Related service

    Automate the boring certainty

    Rules, routing, formatting and repetitive system actions often do not need an LLM at all.

    Use AI where ambiguity lives

    Language, document interpretation, classification and synthesis are stronger candidates.

    Keep approval where risk lives

    Financial, clinical, contractual and high-impact actions deserve explicit human gates.

    Instrument the workflow

    If failure cannot be observed, the automation is not ready to be trusted at scale.

    NOTE 03
    Healthcare AISafetyUX

    Healthcare AI should improve the workflow, not pretend to replace judgment.

    Healthcare is a strong test of product discipline because the cost of confidence without evidence is high. Good systems reduce cognitive load, organize context and make professional review easier.

    Read full field note Related service

    Assist the professional

    Design around the clinician or pharmacist's workflow rather than implying autonomous medical judgment.

    Show the evidence path

    Where possible, users should understand where relevant guidance came from and when it was updated.

    Design for uncertainty

    The interface needs a safe state for insufficient evidence, conflicting information and escalation.

    Measure workflow value

    Time saved, documentation quality, adherence and usability can matter as much as model benchmarks.

    Have a harder question?

    Bring the workflow, data and constraints — not just the tool name.

    Talk architecture