CrewRights AI
An assistant that answers flight attendants' questions about their union contract, used by a working crew member, and built so that a confidently wrong answer is harder to produce than no answer.
Duration
2026
Team
1
Scale
POC
Scope
United States

Overview
CrewRights AI answers questions about the collective bargaining agreement for American Airlines flight attendants, with an assistant called Amelia. I designed and built it as a proof of concept for a working flight attendant, who used it. The code is public on GitHub; the contract text is not stored in the repository.
Challenge
The contract runs to hundreds of pages, and people's words are not the contract's words.
Asking a colleague is fast and often wrong: "unless" is the first thing forgotten.
Asking the union is right but slow, so people guess.
A wrong number about rest, pay or a filing deadline gets acted on, because the reader cannot tell it is wrong.
Strategy
The objective: fast answers a crew member can trust, on a phone, at the moment of decision.
Why now: language models made contract search practical, and also made confident fabrication easy.
The strategic choices: invert the usual goal. Optimise for not being wrong rather than for always answering. Only state values that appear in the cited clause; say plainly when the contract does not cover something; keep citations always visible; force a clarifying question when a clause depends on conditions; treat low confidence as a visible state rather than a refusal. Use a small model deliberately: the rules and the structure carry the reliability, not model size.
How success was measured: real use by a crew member, and answers traceable to a clause every time.
How it works
Intent is classified first at no model cost, so a greeting never triggers a search.
Retrieval is section-aware, and a match expands to its whole clause.
The reasoning step classifies, evaluates and explains whether a clause answers the question.
Seventeen design decisions are recorded with the alternatives that lost.
Status
Proof of concept, built in 2026 and used by a working flight attendant.
Project Artifacts
Project Details
Industry
Airlines
Duration
2026
Team Size
1
Direct Reports
0
Scale
POC
Scope
United States
Budget
Undisclosed
Platforms
Web (Next.js), Supabase with pgvector, OpenAI models
Regulatory
Heavy
Engagement
Proof of concept
