WeatherWear: Smarter travel prep with generative AI on AWS

WeatherWear helps travelers prepare confidently with personalized packing lists and destination guidance across iOS, Android, and the web. The company’s mission is to remove uncertainty from travel preparation by providing accurate weather insights and smart recommendations.

    WeatherWear: Smarter travel prep with generative AI on AWS

    Customer Challenge

    Travelers spent a significant amount of time building packing lists and interpreting airline and destination policies. A rules-based approach struggled with edge cases (multi-city itineraries, family travel, seasonal gear), which hurt first-week retention and increased support volume. WeatherWear sought a production-grade generative AI assistant that could:

    • Generate policy-aware, personalized packing lists from trip context (dates, destinations, traveler profile).
    • Answer travel prep questions with grounded citations.
    • Meet strict latency and cost goals.
    • Operate with built-in privacy, security, and safety controls.

    Why AWS

    AWS provides managed GenAI building blocks, including Amazon Bedrock (models, Agents, Guardrails), as well as retrieval options (Amazon Kendra, Amazon OpenSearch Serverless vectors), and a mature security/ops stack (KMS, CloudTrail, WAF), that accelerates the transition from prototype to secure, scalable production while controlling spend.

    Why Tech Holding

    Tech Holding is an AWS Services Partner experienced in RAG architectures, Bedrock, and mobile backends. The team brought prompt engineering discipline, guardrail design, and a cost-aware operations model, enabling rapid delivery with measurable business outcomes.

    The Solution

    User experience: In-app assistant generates editable packing lists and answers questions (e.g., “Can I bring trekking poles in carry-on?”) with citations to sources.

    Architecture overview

    • Front end: Flutter (iOS/Android), Next.js microsite.
    • API/Orchestration: Amazon API Gateway to AWS Lambda (Node.js) in a multi-AZ VPC.
    • Generative AI:
      • Amazon Bedrock (Claude) for instruction following and tool use via Bedrock Agents.
      • RAG: Amazon Kendra for airline/destination policies and FAQs; Amazon OpenSearch Serverless (vector) for high-churn content (packing templates, brand catalogs).
      • Embeddings: Bedrock embeddings; chunking ~600–900 tokens; metadata (airline, route, season, validity dates).
      • Guardrails: Bedrock Guardrails with deny lists, PII filtering, prompt-injection protections; enforced refusal for medical/visa/legal advice.
    • Data & state: Amazon Aurora Serverless v2 (PostgreSQL) for trips/profiles; Amazon S3 for documents and transcripts; Amazon SQS for incremental re-index and background jobs.
    • Security & networking: Private subnets; VPC endpoints (where supported); AWS WAF on web tier; AWS KMS CMKs per data domain; AWS Secrets Manager; AWS CloudTrail all regions to a protected S3 log bucket.
    • Operations: Amazon CloudWatch metrics/alerts, AWS X-Ray tracing; AWS CodePipeline/CodeBuild/CodeDeploy with canary and automated rollback; feature flags for prompt/model rollout.

    AWS Services Used

    • Amazon Bedrock (Claude, Embeddings, Agents, Guardrails)
    • Amazon OpenSearch Serverless
    • Amazon API Gateway
    • AWS Lambda
    • Amazon Aurora Serverless v2 (PostgreSQL)
    • Amazon S3
    • Amazon SQS
    • AWS WAF
    • AWS KMS
    • AWS CloudTrail
    • AWS Fargate
    • Amazon CloudWatch
    • Amazon VPC
    • Amazon Route 53

    Design choices

    • Model selection: Chosen for long context and strong instruction following relative to lighter models; evaluated latency, cost/token, licensing, and safety surface.
    • Retrieval strategy: Hybrid (Kendra BM25 + OpenSearch k-NN) with re-ranking (light cross-encoder on SageMaker) to boost grounding; metadata filters to constrain airline/route/time validity.
    • Responsible AI: Guardrails, harmful-content blocking, transparent citations, user feedback loop; periodic bias spot-checks across traveler profiles.
    • Resilience & DR: Multi-AZ; DLQs for Lambda/SQS; blue/green deployments; RTO 30 min / RPO 5 min using Aurora PITR and cross-region S3 replication.

    Results and Benefits

    Early outcomes (first four weeks)

    • Users completed 18,900 packing sessions with an 87% task success rate, indicating that most scenarios are handled without human support.
    • The assistant maintained a p95 latency of 1.6 seconds across mobile and web while using hybrid retrieval.
    • A grounded-answer rate of 92% on the weekly evaluation set indicates strong alignment between retrieved sources and generated responses.
    • Cost per task of $0.019 is within the budget envelope, with additional savings planned through prompt compression and context trimming.

    Our Partners

    aws
    Google Cloud Platform
    Cloudflare
    ServiceNow
    Snowflake
    Vanta

    By using this site, you agree to thePrivacy Policy.