Case study 04 / 06
Generative AI · Streaming · Personalization
Travel Itinerary GenAI Sales Accelerator
A GenAI workflow that generates personalized itineraries, maps activities to relevant insurance benefits, enriches them with Google Maps data and streams results progressively over NDJSON.
- Role
- Built the generation workflow and streaming
- Company
- ICICI Lombard
- Stack
- Python
- FastAPI
- LLM
- Semantic retrieval
- Google Maps
- NDJSON streaming
- Next.js
01
Overview
A sales-accelerator experience that generates personalized travel itineraries and links itinerary activities to relevant travel-insurance coverage and add-ons.
02
Context
Long-form generation is slow to arrive in full, and sending a model everything you know is expensive. Both hurt an interactive sales experience.
03
Problem
Generate useful, structured itineraries with less context per request, and show results as they are produced instead of after a long wait.
04
My contribution
- Built a GenAI workflow that generates personalized travel itineraries.
- Worked on activity generation and mapping itinerary activities to relevant insurance benefits and add-ons.
- Used retrieval and context selection to send only relevant information to generation.
- Integrated location and geospatial information with Google Maps.
- Built NDJSON streaming so the frontend renders content progressively.
- Worked on structuring and validating generated output, workflow states and fallback handling.
05
Architecture
Select context → generate → validate → stream
User intent drives semantic selection of only the relevant context. Generation produces structured itinerary items, which are validated and streamed to the client one NDJSON line at a time, so the itinerary builds on screen as it is produced.
- FastAPI service orchestrating the generation workflow.
- Semantic retrieval / context selection to keep prompts small.
- Structured output with validation and fallbacks.
- Google Maps for location data.
- NDJSON streaming to a Next.js client for progressive rendering.
06
Decisions & trade-offs
- NDJSON over waiting for a full response
- Each line is a complete JSON object the client can render immediately, which improves perceived responsiveness.
- Context selection before generation
- Sending less, more relevant context reduced token/memory consumption by roughly 65%.
07
Impact
~65%
Less token/memory use
From context/data optimization
- Streaming improved perceived responsiveness through progressive rendering.
- No conversion or sales-lift figure is claimed.
Described at a public level — no confidential company, customer or system details.