Back to portfolio

Case study

Hubway.ai

Ship an LLM-based virtual assistant across 6 international airports (Paris, Vienna, Belgrade, Dubai, Bastia, Marseille) with a guided, reliable and production-ready user experience.

Company
Hubway.ai
Role
AI Solutions Engineer · Product & End-to-End Delivery
Period
Oct 2025 to Present
Location
Lyon, France
6 airports in productionLLM eval framework ownedClaude Code · Cursor · MCP daily

I drive end-to-end an agentic virtual assistant for 6 airport operators, and I own the LLM evaluation framework (accuracy, hallucination, task success, latency, cost) that drives our release decisions.

The challenge

  • Turn an LLM assistant into something genuinely useful in the field, not a demo chatbot. 6 airports = 6 distinct operational contexts.
  • Combine visual guidance, multilingual answers, conversational context and per-site operational constraints.
  • Give teams a credible evaluation loop to decide what to ship and what to block before production.

What I led

  • Prioritized product roadmap centered on the highest-value passenger journeys and the most critical friction points.
  • LLM evaluation framework (accuracy, hallucination, task success, latency, cost) — owned, drives release decisions and pre-production validation.
  • Daily AI-augmented workflow: Claude Code for shipping, Cursor for code review, MCP for connecting tools. Patterns adopted across the team.
  • Cost & ROI model used in commercial discussions. Architecture documentation, client onboarding, walkthroughs.

Outcomes

  • Production rollout across 6 international airports with a reusable product foundation from one site to the next.
  • Release decisions driven by data via the LLM eval framework, not feeling.
  • Multiplied iteration speed on prompts, visual guidance and multilingual traveler journeys.

Stack & scope

Claude CodeCursorMCPTypeScriptNext.jsAnthropic Claude APIOpenAILLM Evaluation

Tania blends product rigor and execution speed like few engineers I've worked with.

Lead Product, Hubway.ai

Let's talk AI product or agentic systems