Writing from the work
Essays on AI workflows and the future of software.
Thoughts on useful AI, connected knowledge, human judgment, and what engineers need as the work changes.
The representation is not the source
Reliable retrieval preserves a reversible path from every useful representation back to the page a person can inspect.
The user should drive the demo
A working session reveals more when the operator controls the case and the pace.
Start with one thread, not the whole inbox
A narrow slice of real context can test an AI workflow before the product asks for broader access.
A title is only the first retrieval filter
Finding the right person for a workflow requires more than matching a job title to a topic.
The permission list is part of the product
Shipping an AI extension forces its data access and responsibility boundaries into a form users can inspect.
The demo data is already a product decision
Synthetic workflows shape what a prototype can appear to prove, so their assumptions need to stay visible.
An attention agent needs a theory of progress
An attention agent becomes useful when it connects time to the evidence and state changes the work was meant to produce.
Someone has to own the workflow
AI agents become reliable when a named person controls what they may observe, change, and learn from.
Disagreement can be the market map
Conflicting customer interviews become useful when the context behind each claim stays visible.
A pitch deck should show where the evidence stops
A useful pitch deck marks the boundary between current evidence and the next experiment.
A design partner should change the product
A design partnership becomes useful when evidence from a real workflow can change what gets built.
The first course should buy evidence
A bounded experiment can show whether a new field deserves a larger commitment while preserving the skills that already work.
The system of record should stay boring
AI can interpret messy communication while a deterministic system preserves the authoritative state of the work.
AI literacy starts with real work
People understand AI when they can connect its capabilities and limits to work they already know.
A reliable automation keeps a ledger
AI workflows become dependable when every attempt, failure, and safe retry has a visible state.
Access to intelligence begins before the prompt
AI becomes accessible when it can work with the languages, voices, documents, and channels where information first appears.
The template is where judgment begins
Templates encode what repeats, while the exceptions show where people and AI systems still need to reason carefully.
Trust is part of the workflow
Insurance technology becomes useful when it helps people explain decisions and sustain trust across a long relationship.
A merged change is still a hypothesis
AI can accelerate software changes, but the work is only complete when the intended behavior is visible and verified in production.
The interface is where the work already happens
AI workflows become useful when they meet people inside the conversations where work already moves.
AI will eat software one workflow at a time
AI will spread through software as every company looks for places where a system can understand context, take action, and shorten the path to an outcome.
What I want software to give people
The tools I want to build should leave people with more context and more confidence in their own judgment.
What keeps an engineer relevant now
Technical depth still matters, while communication, learning, product judgment, and the ability to make work visible now carry more of an engineer's value.
Human judgment belongs in the product
AI can prepare evidence and reduce coordination while the accountable person keeps a clear place to decide.
A better tool remembers how you got there
Useful memory preserves the sources and changes that shaped a decision, so the next person can pick up the thread.