Consumer AI · 0-to-1 · PRD · LLM Evaluation
An AI packing copilot that thinks like a seasoned traveler — so you don't have to.
Solo PM — Strategy, Research, PRD, GTM
Problem
Travelers over-pack or forget essentials due to generic, static packing lists
Solution
Personalized AI packing recommendations based on trip context, weather, and user history
Stage
Concept to full PRD
Methods
User Personas, Hypothesis Statement, Eval Framework, RAG Architecture, Phased GTM
“Packing lists haven't changed in 30 years. The traveler has.”
Most packing apps give you the same static checklist regardless of whether you're heading to a business summit in Chicago or a surf trip in Costa Rica. The context is always missing.
The real problem isn't that people don't know what to pack — it's that the cognitive load of trip-specific planning falls entirely on the traveler. Every trip starts from zero.
Roadworthy was designed to eliminate that friction by building a packing copilot that learns trip context, reasons about it, and delivers recommendations that feel like advice from someone who's been there.
The users
The Frequent Business Traveler
34, management consultant, flies 3x per month
Packs on autopilot but over-relies on memory, forgets context-specific items like adapters, timezone meds, and client-dinner attire.
Needs: Speed, reliability, and zero surprises on travel day
“I don't need a list. I need a copilot.”
The Intentional Leisure Traveler
29, takes 4–6 trips per year, different types each time
Spends too long researching what to pack for new destinations or activity types.
Needs: Contextual guidance without having to start from scratch every time
“Every trip feels like the first time I've ever packed.”
My approach
01
Wrote a hypothesis statement anchored to a specific, measurable user outcome. Framed the core assumption: if Roadworthy delivers contextually accurate recommendations on the first attempt, users will skip manual list-building entirely.
02
Audited existing packing apps including PackPoint, TripList, and Google Trips. Identified the gap: none use trip context dynamically. All are static or semi-static list tools.
03
Defined the RAG architecture approach: retrieve relevant packing data and augment LLM prompts with destination, weather, activity type, and past-trip context before generation. Wrote prompt constraints to reduce hallucination and enforce category-based output structure.
04
Produced a full PRD covering problem framing, personas, functional requirements, model evaluation criteria, eval framework, edge case handling, and a phased GTM rollout plan.
The PRD highlights
Defined success criteria for AI output quality: accuracy, relevance, completeness, and tone. Built a scoring rubric for human eval and automated regression testing.
Established hard constraints to prevent the model from generating unsafe, irrelevant, or over-generic recommendations. Defined output schema: categories, quantities, priority flags.
Phase 1: iOS MVP with manual trip input and AI list generation. Phase 2: Weather and calendar API integrations. Phase 3: Learning loop — user edits feed back into personalization layer.
Hypothesis statement
“We believe that frequent travelers who struggle with context-specific packing decisions will successfully generate a complete, trip-accurate packing list on their first Roadworthy session — without editing more than 20% of AI-generated items — because the RAG-augmented prompt delivers contextual specificity that static list tools cannot match.
We will know this is true when 70% of beta users complete their first packing list without requesting a regeneration.”
What I learned
The most important product decision on Roadworthy wasn't the feature set — it was deciding what the AI should never do. Prompt constraints and eval frameworks aren't just engineering concerns. They're core PM responsibilities on any AI product.
Defining the hypothesis before building the PRD kept every feature decision anchored to a real user outcome. It also made trade-off conversations easier — if a feature didn't move the hypothesis metric, it didn't make Phase 1.
This project sharpened my conviction that great AI product management is less about prompting the model and more about designing the system around it.