OO
←Back to Work
FlutterAI-Powered ProductsMobile Performance

Packy AI

Sole mobile engineer — independent build, owned the entire client from architecture to release.

Combining AI-generated travel recommendations with deterministic user controls - keeping suggestions explainable and latency low even when model confidence varied.

Packy AI hero screenshot

01 —

The Problem

Combining AI-generated travel recommendations with deterministic user controls - keeping suggestions explainable and latency low even when model confidence varied.

02 —

Approach & Architecture

Delivered a planning workflow that generates practical travel itineraries using preferences, trip context, and user constraints. The implementation balanced AI-generated suggestions with deterministic interaction patterns to keep results explainable and reliable. The product focused on maintaining low-latency responses and preserving usability even when recommendation confidence varied.

Fast iteration mattered specifically here — the recommendation UI went through many rounds of tuning against real model output, and reshaping the client quickly without a native release cycle for every experiment kept that iteration fast.

Key decisions

  • Combined deterministic client orchestration with AI-backed recommendation services.
  • Normalized AI responses into stable internal view models for predictable rendering.
  • Added fallback and re-ranking paths when generated output did not meet quality thresholds.

Security considerations

  • Applied input sanitization and prompt guardrails for model interaction boundaries.
  • Constrained profile data usage and removed non-essential payload fields in recommendation requests.
  • Optimized response latency with progressive rendering and cached itinerary fragments.
Packy AI screenshot

03 —

Technical Challenges

Integrating AI into a mobile experience isn't just an API call — it's a latency problem. The real challenge wasn't calling the model; it was designing an experience where the user always understood what the app was doing while it waited on a response.

04 —

Trade-offs

I deliberately kept some AI-driven features simpler than they could have been in the first version — instead of making every interaction fully dynamic, I focused on predictable, reliable results with clear loading and fallback states.

05 —

Outcome

Live on App Store and Play Store. Delivers personalized travel itineraries with fallback reliability across varying model output quality.

06 —

Lessons Learned

AI should improve the experience, not dominate it. Users care far more about responsiveness, clarity, and reliability than whether a feature is powered by the most advanced model available.

More Screens

Packy AI screenshot 2