JourneyMind AI

Background

In travel, disruption decisions carry cost and contractual weight. Pure AI answers are not enough because they cannot be audited. I wanted to show that AI orchestration can be combined with a deterministic, explainable decision core — even under hackathon time constraints.

Problem

When travel is disrupted, operators must reconcile passenger context, fare rules, and rebooking options quickly. Manual handling is slow and inconsistent; a purely generative AI approach is unaccountable and cannot be audited.

Goals

• Use AI for context and option preparation, not the final decision
• A deterministic, rule-based decision engine with rationale attached
• REST API between React frontend and FastAPI backend for easy integration
• End-to-end tests of the primary operator flow with Playwright

Architecture

Technology stack

React · TypeScript · Vite · Tailwind CSS · FastAPI · Python · REST APIs · Playwright

Engineering challenges

• Separating probabilistic AI orchestration from deterministic decision-making
• Making rationale a true output of the decision engine, not generated afterwards
• Modelling constraints so they remain machine-evaluable and human-readable
• Hackathon scope: complete architecture, but deliberately not all production concerns implemented

Trade-offs

• A deterministic core provides auditability but less flexibility than an end-to-end model — the right trade for a regulated domain
• REST instead of GraphQL/gRPC: a small, inspectable surface with no tooling overhead; transport is not the bottleneck here
• Explainability requires more data plumbing, but it is the main value of the system
• Hackathon scope: architecture and core flows are complete; auth, load testing, and live data integration are designed for but not implemented

Lessons learned

• Determinism alone does not guarantee explainability — rationale must originate in the decision process
• Keeping AI orchestration outside the final decision step preserves auditability
• End-to-end tests of the operator flow reveal more than isolated unit tests
• Explicit constraints are more valuable than implicit model flexibility in regulated contexts

Current status

Hackathon build with complete architecture and core flows — no live data integration or production hardening.

Future improvements

• Production hardening: auth, rate limiting, observability, resilience
• Live data integration for disruptions, fares, and rebookings
• Configurable constraints without code changes
• Audit log for recommendations and manual overrides

Estimated milestones

• Q3 2026 — Auth and basic observability
• Q4 2026 — Configurable constraints and audit log
• Q1 2027 — Live data integration and load testing