Ask Markerr: Designing an AI Rental Search Experience for People Who Want to Trust It, But Shouldn't Blindly
Ask Markerr was a strategic prototype designed to help Markerr move from a data platform into a credible AI product company. I used AI-assisted desk research, behavioural design, and trust-centred interaction patterns to shape a conversational rental-search prototype for people who want AI to understand them, but still need proof before trusting it.
A note on honesty, up front
Ask Markerr did not ship to real users. It reached POC/MVP stage and was demonstrated to partners and stakeholders as part of Markerr's AI roadmap. That distinction matters, especially in AI product design, where the gap between demoed and validated can be large.
01 // Architecture Matrix
Beyond the chatbox.
Ask Markerr was not a chatbot veneer over a real-estate database. It was a strategic POC/MVP prototype for a harder question: how do you let renters use human language without asking them to blindly trust fluent AI output?
Scope Contract
Starter prompts teach what the assistant can and cannot do before the first generation.
Semantic Reflection
The assistant mirrors parsed intent before resolving listings.
Grounded Output
LLM text maps into deterministic property cards rather than inventing facts.
Failure Recovery
Correction controls sit inside the conversation stream instead of hiding in support states.
02 // Core Design System
Trust comes from structure.
My design principle was simple: the assistant should feel human enough to approach, but honest enough to check. Every pattern had to make interpretation visible, evidence inspectable, or mistakes easy to correct.
Input
Starter scenarios convert life context into clearer intent.
Reasoning
Streaming states expose what the system is checking.
Output
Structured cards keep facts anchored to database objects.
Standard UX Pattern: High Cognitive Anxiety.
The user sees delay without knowing whether the system is stuck, reasoning, or silently dropping a constraint.
Stakes-Calibrated Friction Layer.
Waiting becomes inspectable: the interface shows what is being parsed before asking the renter to trust a result.
03 // Trust Mechanics
Design for the moment AI is wrong.
The central failure mode of conversational rental search is semantic drift: a constraint disappears, but the answer still sounds confident. I treated correction as a first-class interaction, not an apology after the model breaks.
Design Decision 7
Degraded correction flows
Users can re-steer specific context tokens without wiping the memory block, restarting the search, or wondering whether the assistant remembered the wrong thing.
User Prompt
Find a 2-bed rental under $3,200 near transit.
I am moving in three weeks and have a Golden Retriever.
Short commute matters more than square footage.
The model returns a confident match, but the pet-policy constraint silently drops from the parsed context.
Inline Correction Layer
Parsed constraints: under $3,200, near transit, 2 bedrooms,Allows Golden Retriever, commute priority.
04 // System Telemetry
Measure trust, not applause.
Because this was a prototype, I framed success as a behavioral instrumentation plan. A future live release should measure where users verify, correct, abandon, or over-trust the assistant.
Trust in generative AI systems is earned by making interpretation visible, evidence inspectable, and inevitable mistakes simple to correct.
Continue through adjacent product systems.
Designing trust into a government-scale AI audit system
AI-assisted legal instruction authoring for the Brazilian Federal Court of Accounts. Writing time reduced by 63%.
Two platforms, one product: B2B pensions for HR and brokers
B2B platform architecture and multi-surface workflow design for HR managers and brokers.
Ditching Spreadsheets and Empowering Financial Plans
Fintech workflow separation and analytical surface strategy. Replacing spreadsheet-as-product behavior.