From Information to Insight: Redefining Travel Planning using Alexa's AI
TL;DR
I led the 0-to-1 design of Alexa's first generative AI travel experience: an LLM orchestration framework that turns an ambiguous voice request into a structured, personalized itinerary. The patterns and principles I established now serve as the reference architecture for all future Alexa expert advisor experiences.
My role: Lead UX designer - interaction model, LLM orchestration, cross-org alignment across 8+ teams Outcome: 88% positive sentiment · 6.9/7 CSAT · 96% API accuracy at launch
The problem
Travel planning is broken - the average traveler views 141 pages before making a booking decision - manually synthesizing reviews, guides, and booking sites into something coherent. 55% of them start without even knowing where they want to go. Existing tools are reactive and generic. Nobody had built a voice-first, LLM-powered travel advisor that could handle the ambiguous, top-of-funnel moment where the real planning begins. That's what I set out to design.
Goal : Leverage Alexa's unique capabilities (voice accessibility and LLM power) and expert travel content from 3P partners to provide an experience that focuses on the ambiguous, top-of-funnel discovery and hands-free iteration.
The solution
Role and Scope
Lead UX Designer
Established the interaction model and design patterns for a category that had zero prior design or engineering reference points.
Navigated extreme ambiguity by leading the cross-functional effort to define how to effectively integrate third-party expert content via proprietary APIs and an internal LLM, moving from raw data to a conversational, trip-planning utility.
Managed risk via the 'Crawl, Walk, Run' framework.
Duration
~ 8 months
Target devices
Echo devices, Web, Mobile, FireTV
Team collaboration
Cross functional: Product management, engineering, QA, Business Development, Marketing ;
Cross organizational: Local Info, Alexa web, Alexa mobile, Device & Design Group(visual design team), Multimodal developer tools, Smart briefings
3P content partners
Partners considered: Fodors, Trip Advisor, Lonely Planet, Travel & Leisure, US News & travel, GetYourGuide
Impact
This experience was released in Feb 2025 for selected participants as a part of the new Alexa+, in partnership with Fodor’s travel.
“Considering Alexa’s convenience ... and its integration with Fodor’s and TripAdvisor, Alexa+ has its place for top-of-the-funnel research.”
6.9/7.0
Strong customer satisfaction scores from Beta cohort
88%
Positive sentiment from the early Alexa+ beta cohort
2.66% CPDR
Customer Perceived Defect Rate- exceeding the 15% launch target.
96% LLM Accuracy
API Invocation and expert-selection accuracy
Critical Decisions & Collaborations
-
The Challenge: The newly deployed LLM model struggled with the highly specific tone, factual requirements, and conversational flow necessary for trusted travel advice. The model needed intensive grounding to move from general knowledge to specialized, reliable "Expert Advisor" output.
The Decision: I led the collaboration with Engineering to perform iterative model tuning by defining and writing conversational exemplars (high-quality, desired input-output pairs). This involved:
Model Testing: Actively testing the experience with different LLM model versions and reporting design-critical findings
Prompt Engineering: Working closely with Product and Engineering to strategically refine the instruction prompts and parameters to resolve specific design issues (like over-verbosity or lack of structure).
Impact: This hands-on, iterative work directly controlled the quality and behavior of the LLM. It was critical for achieving the necessary conversational fidelity required for a good user experience.
-
The Challenge: Establishing the boundary between the existing 1P Alexa information services (the "first line" of defense) and the new Expert Advisor experience to ensure efficient routing, prevent redundancy, and provide most relevant information to the user.
The Decision: I devised and championed the Intent Routing Strategy with the 1P Info Org. This established a clear decision tree for the Alexa Intent Router:
Info Org (1P): Handles simple, factual, or singular requests (e.g., "What is the capital of France?").
Expert Advisor (LLM): Handles complex, multi-faceted, or generative requests (e.g., "Plan me a five-day itinerary for France that avoids crowds and is good for kids.").
Impact: This collaboration was crucial for the long-term adoption and architectural sustainability of the Expert Advisor platform.
-
The Challenge: We faced performance challenges rooted in the novelty of the platform: 1) High API and LLM Latency 2) The necessity for real-time responsiveness in a VUI; and 3) A Small Initial Image Database which led to inadequate visual search results and poor user experience on screen-enabled devices.
The Decision: I led the partnership with Engineering to address these issues holistically, prioritizing experience reliability over initial feature richness. Our solution involved
Latency Masking Strategy: Designing and implementing predictive visual updates and conversational fillers to actively mask the inherent latency, ensuring the user perceived a responsive, engaged system (Tenet 3: Graceful Failure).
Fallback Design: Championing image fallback mechanisms when high-fidelity visual assets were unavailable, ensuring the user could always complete their task.
Impact: This integrated approach ensured we met the aggressive launch performance budget and mitigated the risks associated with immature technology, protecting the user experience
-
The Challenge: We faced a challenge of fighting the LLM's natural verbosity while needing to standardize the display of complex information (itineraries, POIs) for a new multi-modal category.
The Decision: I worked with design patterns team to define certain patterns like:
Data Summarization: Defining strict summarization thresholds
Visual Display: Creating the Card Pattern for screen-enabled devices to standardize the display of POIs and itinerary segments for scannability.
POI Detailing: Designing separate concise spoken summaries and detailed visual information, optimizing content for multi-modal consumption.
Impact: This work established the foundation for all subsequent Expert Advisor categories, significantly reducing design iteration time and technical debt for future teams.
Reflections
Key Learnings
Defining the AI framework: I learned how to bridge the gap between prompt engineering and UX. Understanding LLM limitations allowed me to architect a framework that translated raw partner data into trustworthy, conversational advice
Strategic cross-team orchestration: In a complex organization, "seams" in the user experience are usually a result of team silos. I learned to lead by defining clear technical handoffs and scope boundaries, ensuring that regardless of which team built a feature, the user experienced a single, unified Alexa.
What I’d do Differently
Parallel Pathing Technical Spikes: I would have advocated for deeper "technical spikes" (early engineering explorations) alongside the high-fidelity prototypes for partner pitches. This would have narrowed the gap between the "vision" sold to partners and the initial "Crawl" phase of the LLM’s actual performance.
Accelerated Cross-Org Strategic Alignment: I would have initiated formal syncs with the Alexa Information Organization much earlier in the discovery phase. Reaching a definitive conclusion on scope boundaries and data ownership sooner would have reduced redundant efforts and allowed us to focus on the unique conversational logic of the Expert Advisor rather than negotiating baseline information retrieval.
What didn’t go as per Plan
The Problem: The "Vision-to-Reality" Gap
What Happened: I designed high-fidelity ideal visions that showcased a perfect, seamless AI conversation. However, the internal LLM was still evolving, and early versions couldn't always fulfill the complex multi-turn logic that would add the delight factor for the users.
The Pivot: I pivoted from "designing the ideal" to progressively disclosing complexity as tech evolves.
The Lesson: When designing for emerging tech, the UX must be as flexible as the API. Ownership means being responsible for the "worst-case scenario" just as much as the "happy path."
Process Deep-Dive
Defining the Problem
Desk Research
Travel planning is often full of information overload and friction. The average traveler views 141 pages of content before making a booking decision. This fragmented experience presented two key problems:
High Cognitive Load: The user has to manually synthesize disparate pieces of information (reviews, guides, booking sites) into a coherent trip plan.
Lack of Personalization: With 55% of travelers starting without a specific destination in mind, the existing tools are reactive and non-customized, failing to serve the user who needs inspiration and personalized direction. 57% of US travelers feel that brands need to tailor information based on personal preferences
Audit - 3P Partner Websites
Goals of the audit
To understand what the key value props are
To understand what kind of information are users going to these websites for
Key insights
Travel websites offer travel inspiration ideas, travel itineraries, things to do, restaurant and hotel recommendations, and other planning tools
They generally ask the user for key pieces of information like destination, time/duration of travel, travel preferences, budget, etc.
Design Considerations
Understanding the capabilities of LLMs: Partnered with product and engineering to define what Alexa’s LLM could realistically do- and how to design for a non-deterministic system.
Understanding the key value propositions of Alexa: Clarified Alexa’s unique value in the travel journey, identifying top-of-funnel discovery and hands-free iteration as the strongest use cases. And intentionally de-scoping transactional features for P0.
Foundational Design Tenets
Strategic Goal - Redefined
I reframed the goal around making Alexa a useful travel advisor by leveraging its strengths- voice accessibility and LLM generation- to replace today’s fragmented planning experience. This required architecting a system that could:
Synthesize disparate 3P content into coherent itineraries.
Contextualize recommendations using user history and conversational context.
Maintain continuity across Echo, Web, and Mobile by choosing the right modality for each task.
Foundational Design Tenets
Build the Platform, Not Just the Product : Create flexible interaction patterns and reusable multimodal standards (including pattern structure, knowledge summarization methods, and error handling flows) to support all future Expert Advisor categories.
Prioritize User Utility Over Platform Preference: The system must relentlessly optimize for the user's best path to completion, regardless of the originating device, content source (1P vs. 3P), or modality.
Lead the Conversation, Deliver Concisely: Keep responses proactive, structured, and succinct to help users navigate complex planning.
Design for Conversational Graceful Failure: The system must never dead-end or sound robotic. Instead, it must apologize, pivot to a known capability, or offer an intelligent alternative path to continue the user's planning session.
Conceptual Design & System Architecture
Data modeling - architecting the LLM orchestration layer
Designed the core architecture that transforms ambiguous user input into reliable, actionable travel guidance. This framework defined the boundaries and requirements for AI Science and Engineering and became the foundation for all Expert Advisor functionality. The three core architectural elements were System Inputs (The context), Orchestration layer (The logic) and System outputs (The utility).
-
1. System Inputs (The Context)
Inputs necessary for the LLM to deliver personalized and context-aware responses.
Ingress utterances
Conversational History and Follow-up Questions
Personalization Signals (e.g., historical user preferences)
Device Types (e.g., Echo, mobile, screen-enabled)
2. Orchestration Layer (The Logic)
Core functional architecture to process inputs into controlled outputs.
Alexa LLM Layer: Controlled the tone, summarization and top-level actionable next steps. (per Tenet 3: Lead the Conversation, Deliver Concisely)
Travel Partner Content Layer: Provided the essential knowledge bases, structured content, and human-written exemplars used to ground the LLM's responses, ensuring accuracy and mitigating hallucination risks.
Business Logic Layer: Housed the prompts and algorithms for generative tasks, such as creating itineraries and requesting the appropriate data from APIs.
3. System Outputs (The Utility)
Structured Output: What Alexa would deliver (e.g., itinerary objects, points of interest).
Multi-Modal Delivery: Modality for each output component (e.g., voice for summary, screen for detail, mobile for continuity hand-off, per Tenet 2: Prioritize User Utility).
Day in the Life - From Ambiguity to Itinerary
This validated the LLM Orchestration Architecture and translated abstract strategy (data models, flows) into concrete, high-value interactions.. It enabled the entire cross-functional team to literally walk in the user's shoes over the span of weeks and months. There were 4 main phases designed, along with ingresses, key value props, and modality.
The key outcomes driven by this artifact were:
Identifying Key Value Propositions: Validated that proactivity, personalization, and cross-device interactions were integral to user adoption
Defining Example Use Cases: Established feature sets based on realistic user needs, which drove the Crawl, Walk, Run strategy.
Defining Ingresses: Ensured a robust, non-linear entry and exit strategy
Identifying Data Sources & Devices: Provided the definitive requirements matrix for the Engineering team, clearly linking user actions to the necessary data source (1P, 3P) and the optimal device for fulfillment.
The Vision
North Star vision: A fully realized Expert Advisor that goes beyond basic trip planning to become a proactive, personalized, multimodal travel concierge. Some of these capabilities were intentionally deferred from P0 but shaped the long-term architecture.
Invisible Modality
Alexa intelligently delivers information on the most effective device—voice for quick changes, web/mobile for complex decisions, and mobile for real-time reminders- without the user needing to specify the modality.
Predictive Personalization
Alexa autonomously generates tailored itineraries using past behavior, preferences, and calendar context, guiding users from vague ideas to actionable plans with minimal prompting.
Interactive Multi-Modal Summaries
Rich, touch- and voice-driven itinerary components enable quick scanning and deep dives, solving information-density challenges on Echo Show while enhancing the conversation.
Partner-Agnostic Recommendations
A dynamic attribution framework ensures Alexa surfaces the best 1P/3P content with proper credit, balancing business requirements with conversational flow and user trust.
What Got Shipped: The Realistic MVP
Launching an LLM-based product with new APIs and third-party content required a disciplined Crawl-Walk-Run approach. I strategically narrowed P0 to only high-confidence LLM tasks and designed the architecture to scale to future complexity without rework. This ensured a reliable launch, protected the team from scope creep, and generated the data needed for later “Walk” and “Run” releases.
Key Features Launched
Discovery & Ideation: High-level, voice-first travel brainstorming and destination recommendations powered by the LLM, but constrained to the knowledge of our single 3P partner.
Itinerary generation: The core capability to generate a multi-day trip plan based on a user's initial high-level criteria (e.g., location, duration, interest).
POI details: The ability to retrieve and display Point of Interest (POI) details with a focus on quick utility. This involved synthesizing 3P data into a concise spoken summary and a scannable visual card that prioritized key information (e.g., hours, location, contact) for decision-making.
Cross-Device Continuity: While the functionality was optimized for Echo devices, there was responsiveness built into each component so as to be scalable on all the other modalities too (including defining VUI for screenless devices). There were also features built in like "Send to phone" to establish the critical first layer of the Human-AI Handoff for complex tasks.
Key Decisions
Scoped P0 to reliable use cases- broad discovery and simple itineraries only; deferred high-risk comparisons to avoid hallucination.
Defined a strict, latency-resilient feature set to maintain speed and responsiveness despite early infrastructure constraints.
Launched with one 3P partner to simplify attribution, reduce legal/engineering overhead, and de-risk the timeline.
The resulting Minimal Lovable Product (MLP) was a highly reliable, focused experience that proved the core value proposition while strictly adhering to the technical and business constraints of the Crawl phase.
Design Validation
The designs were put through Rapid Iterative Testing and Evaluation (or RITE testing) and had these key outcomes:
Participants want to refer multiple sources for their trip planning
Participants would like Alexa to ask before making any changes to the itinerary
Participants would like access to this information across modalities
Usability feedback
Need for high information density
POI detail page - Tour recommendations were not found to be very useful, and they’d prefer other information like history, map, etc.
Scaling Beyond ..
Evolving the Vision: From MVP to Travel Concierge
Next up will be accelerating utility and adoption based on real production data. The original North Star Vision remains the goal, but the path is now clearer, focusing on three core strategic shifts:
1. From "One-Stop Shop" to "Seamless Orchestrator"
I do not believe Alexa should strive to be the "one-stop shop" for every single task. Instead, the vision must evolve into Alexa becoming the seamless orchestrator of the planning journey.
2. Evolving Voice-First to "Ambient Intelligence"
We must evolve from reactive voice-first to a truly ambient, context-aware service. This means the service should be designed to run in the background, only surfacing information via voice, display, or notification when it is proactively helpful- for example, spotting a flight deal related to a destination recently researched via voice, and proactively notifying the user on the kitchen Echo Show.
3. Deepening Personalization via Saved Context
The next phase must focus on integrating more personalization signals like Calendar data (e.g., booked vacation days) This strategic use of personalization will move the experience from being merely helpful to being indispensable.
Framework for Content-Based “Expert Advisor” Experiences
The Travel Planning Expert Advisor was not just a successful launch; it was the foundational reference model for all subsequent LLM-powered content experiences. My work culminated in the creation of a reusable strategic playbook that templatized the complex process of designing and launching this new product category.