The constraint

Travellers do not want a chatbot. They want an answer about the day they are actually having — which means grounding in live state: the itinerary, where they are standing, how long it really takes to get somewhere, the weather on the way. And it means voice, which is unforgiving. A conversation dies above a few hundred milliseconds of awkwardness, so every enrichment had to fit inside a real-time budget. Grounding freshness and turn latency traded against each other explicitly, rather than by accident.

What I built

Co-founded with a partner who leads go-to-market; we work the product together near-daily. Engineering, product design, brand and the infrastructure are mine.

• A LiveKit voice concierge on k3s, grounded in live itinerary, location, travel-time and weather context.
• The retrieval layer behind it: PDFs, web pages, documents and scanned images through OCR, chunking, dedupe and contextual enrichment into Qdrant; query expansion, reranking and reciprocal-rank fusion on the way out. Exposed as agent tools, so voice and text share one knowledge base rather than drifting apart.
• Hybrid search over deals and business profiles in Postgres — vector, full-text, rating and recency together — so the agent can tell a strong match from a merely nearby one.
• One product across three surfaces: a React Native app, a web client that mirrors it, an embeddable widget, a partner portal for hospitality operators, and affiliate attribution built on a probabilistic fingerprint plus install-referrer.

Decisions worth defending

• Grounding beats eloquence. Users forgive a plain sentence. They do not forgive being sent to a closed restaurant across town — so the latency budget went to freshness rather than phrasing.
• Real distance, not naive maths. Straight-line distance is confidently wrong about islands, one-way roads and water, and it routes people wrong while sounding certain. Travel time is computed, not approximated.
• Hybrid retrieval rather than vector-only. Semantic similarity alone will happily return somewhere plausible, closed and badly rated. Full-text, rating and recency are what separate a recommendation from a guess.
• Voice as a performance audit. It surfaces every latency sin in the backend, because a person waiting to be answered notices what a spinner hides.

Operating it

The agent is the smaller half of the work. Paid users were once not credited because webhooks went missing in flight — money and entitlement disagreed, and the app was the last to find out. What followed is the part that makes it a product rather than a demo: signature verification on every event, server-side reconciliation, an admin make-whole API for when the two ledgers diverge, a health check every fifteen minutes that alerts the moment the purchase-to-grant path degrades, and a written postmortem with replay tooling. Store operations, monitoring and incident response are the rest of it.

Where it stands

Live in the App Store and Google Play, launched Hawai‘i-first, and early. The growth numbers that would turn this into a different kind of story are not mine to quote yet.