How to build a chatbot in 2026,
for now

The problem. Booking travel as a new parent is hard. I don't have time to research everything, and I don't know what I don't know — lap infant rules, stroller policies, which terminals have the right connections, how to use our miles, whether non-stop is actually worth it. Generic search gives me 40 tabs and no answers.

The solution. A single assistant that already knows our family, already knows the sources worth trusting, and can give me one clear answer instead of a pile of options to sort through. That's Alon-Ze par Sebkhi — a personal travel assistant built for how we actually travel.

• 1
  Child-first routingNon-stop is the default — not because I had to ask for it, but because I designed it knowing Elio. A 3-hour layover with a lap infant isn't a travel option; it's a stress event.
• 2
  The $300 hard ruleIf an alternate airport saves $300+, it tells me proactively. I built that rule in — no need to ask.
• 3
  Warm directnessOne clear recommendation, brief rationale, concise bullets. New parents don't have bandwidth for decision fatigue.
• 4
  Only sources I already trustEvery source in this system is one I'd actually reference myself. AFKL Open API for live flight data. Official Air France policy pages. The ATL and Paris Aéroport websites. And Rick Steves — because his editorial voice is practical, honest, and earned over decades. No SEO farms, no aggregator noise. If I wouldn't use it to plan my own trip, it's not in the bot.
• 5
  Temperature calibrated at 0.4–0.7Warm in tone, precise on facts. Confidence indicators surface when the bot is hedging — no false certainty.

I built it in three parts: a chatbot brain that speaks in full sentences, a live connection to Air France flight data, and a curated library of trusted sources it can reference. Here's what's actually under the hood:

I designed this as more than a chatbot — it's a small, continuous research practice. Three feedback layers work together to keep the experience honest and improving:

• Qualitative — what users say After each conversation, users can rate the session and leave a comment. This captures the nuance that numbers miss — the difference between "it gave me an answer" and "it actually helped."
• Quantitative — what the scores show Every rated session feeds a live dashboard: average scores, score distribution by session, and flagged low-confidence responses. Patterns surface automatically.
• AI council — background quality check A second AI layer evaluates each response against the original prompt and source material. It flags answers that feel off, hedged, or unsupported — a built-in editor running quietly in the background.

Auto-improving, but never autopilot. The system surfaces what needs attention — but I decide what to update, rewrite, or leave alone. The goal is a lightweight but real research practice: signals coming in continuously, me steering the decisions.

The real test is whether it works in the wild — and it did. I used it to plan our family trip to France: non-stop routing, lap infant logistics, Flying Blue miles, the whole thing. Then I shared it with my sister, who used it to plan her own trip — her, her husband, and their baby, also flying out of Atlanta. Two families. One assistant. Zero generic search results.

That's the difference between a demo and a tool people actually use.