Josh Estrada
Back to work
Enterprise SaaS · Sole Product Designer

AI Assisted Maintenance Troubleshooting

Designing an AI conversation flow that helps residents troubleshoot maintenance issues themselves

Product UI is proprietary. Diagrams represent the design thinking, not the actual interface.

Conversation Architecture

Designing the conversation, not the screen

How AI-guided follow-ups and troubleshooting replace manual back-and-forth. The conversation flow is what I designed and shipped.

Vague Request
"my sink is leaking"
"ac not working right"
"something smells weird"
No context, no troubleshooting,
no priority signal
ai processes
Conversation Architecture
Follow-up Questions
multi-picker · gathering breadth
Troubleshooting Steps
single-select · constrained action
Escape Hatches
skip bail out
emergency detection active
resolves
Dual Outcome
Self-Resolved
Issue fixed immediately.
PMC gets a record, no action needed.
— or —
Enriched Request
Forwarded with full context.
PMC gets richer data upfront.
Better outcome either way —
fewer back-and-forth cycles
Rescued from a shelved project

Originally shelved by the web team. Reframed as a mobile-first initiative and led it after recognizing the opportunity.

Input modality matched to cognitive task

Multi-picker for exploratory follow-ups (gathering breadth). Single-select for troubleshooting steps (constrained action). The input type changes because the cognitive task changes.

Every step skippable

Residents can skip any AI question or bail out entirely. The AI watches for discomfort signals and respects them. No dead ends.

PMC Validation

PMCs confirmed the AI asks exactly the questions they would ask manually, which meant the conversation logic was ready for real tickets

Team Adoption

The web design team started using the same working-prototype approach to test their own AI features

Engineering Blueprint

The system prompt I wrote during prototyping was used as engineering's production starting point

Component anatomy

Why the interaction pattern changes at every step

This looks like a chat interface, but the underlying flow is structured. Each interaction pattern was chosen to match what the resident is trying to do at that point in the conversation.

AI Follow-up
Which areas are affected?
"What else do I need to know?"
AI BEHAVIOR
System prompt generates contextual follow-ups

The AI generates contextual follow-ups based on the maintenance description. Multi-picker lets residents select multiple applicable answers, which fits exploratory questions where breadth matters more than precision.

Why multi-select At this stage, residents are still figuring out what is relevant. The AI lets them tag everything that applies rather than forcing a single answer.
Try this
Turn off the water valve under the sink and check if the leak stops.
This worked Still leaking I can't do this
"Can I fix this myself?"
AI BEHAVIOR
Decision tree narrows with each response

When the AI has enough context, it provides the most likely troubleshooting step. Single-select constrains the response — the resident either confirms success, reports failure, or skips. Each response narrows the AI's decision tree.

Why single-select The resident is doing something physical and reporting back. Binary or ternary choices match that kind of task.
Emergency Detected
This sounds like it could be a gas leak.
Contact emergency services immediately
911
This is not an emergency
"Is anyone in danger?"
SAFETY SYSTEM
Always-on monitoring layer

The AI monitors every response for emergency signals — gas leaks, flooding, electrical hazards, anything involving immediate risk. It runs throughout the entire interaction, on every message.

Always-on monitoring Emergency detection runs as a continuous layer, not a discrete step. The moment risk is detected, the UI changes completely.