Mustapha Adubiagbe

AI accelerated the loop. It did not own the judgment.

I used AI as a pair programmer for alternatives, scaffolding, test ideas, and editing. Product constraints, architecture, review, device diagnosis, and evidence remained human-owned.

AI helped with

  • Comparing implementation approaches
  • Accelerating component and test scaffolding
  • Proposing duplicate and out-of-order failure cases
  • Auditing accessibility semantics
  • Editing documentation after decisions were made

I remained accountable for

  • Choosing the second-screen product problem
  • Defining the recovery contract
  • Selecting the architecture and visual direction
  • Reviewing every code change
  • Diagnosing physical-device failures
  • Deciding which performance claims were supportable

Why this distinction matters

The first “working” implementation failed on a real phone.

The haptic was not perceptible. VoiceOver did not announce the goal. Reduce Motion still allowed the card pulse.

Device testing, not generated confidence, changed the implementation: heavy impact haptics, queued high-priority announcements, and a runtime Reduce Motion check immediately before animation.

AI shortened the iteration. Physical evidence closed it.

Five guardrails

  1. 01

    Every behavioral claim needs a test or a recorded manual check.

  2. 02

    AI output remains a proposal until it fits the domain model.

  3. 03

    Generated copy cannot overstate the benchmark environment.

  4. 04

    Failures stay visible; code does not manufacture success-shaped fallbacks.

  5. 05

    The repository preserves tradeoffs, limitations, and validation commands.

The result

AI increased the number of ideas and checks I could evaluate in the available time. The engineering value still came from selecting, integrating, testing, and sometimes rejecting those proposals.

That is the workflow I would bring to a team: faster exploration, explicit ownership, and evidence before confidence.