Armaan Rana

A travel agent types a raw trip request and gets a bookable quote ready to customise in under 2 minutes, replacing a four-hour manual job. The LLM interprets the request, solvers and live inventory do the planning, staged validation grounds the output, and every agent edit feeds back in. Nine of ten quotes go out unedited.

A brief arrives as prose: three cities on one line, a date range given as “last week of Feb”, a hotel named the way the agent remembers it rather than the way any supplier lists it. The naive answer is to have a model write the itinerary, which produces confident, plausible, unbookable output. The work is deciding which parts of the job a language model should touch at all, and grounding everything it produces against inventory that changes underneath you. The model is load-bearing exactly once, turning prose into a structured request. Everything downstream is deterministic, checkable, and able to refuse.

Reads buying intent signals to set discount depth and find the markup levels where bookings convert. Built for flights and hotels, deployed across hundreds of partner agencies.

A single markup is wrong in both directions at once: high enough to lose the price-sensitive segment, low enough to give away margin on agents who would have booked anyway. So the problem is inference rather than arithmetic — estimating how much buying intent is present in a given request, then finding the markup level at which that segment still converts.

Built TravClan's AI Ops function from scratch: defined its scope, hired its first engineers, and took AI agents into production that now close flights operations and customer support tickets end to end.

Operations work is a long tail. A few ticket types carry most of the volume and the rest are individually rare but collectively large. Closing one end to end, rather than drafting a reply for a human to send, means the agent has to recognise when it is out of its depth and hand back — a boundary conservative enough to trust and loose enough to be worth having.

Shipped the post-booking operations pipeline, live across 20+ destinations, cutting around 50 hours of manual ops work a week.

Confirming a booking is where the operational work starts, not where it ends. What follows is a sequence that varies by destination, is individually simple, and is collectively expensive because a person has to remember every step and notice when one has not happened. Twenty-plus destinations do not share one process, so the variation has to be configured rather than coded.

Live on partner support and sales calls, alongside the chat automation layer.

Voice is the least forgiving channel to automate. There is no scrollback, no chance to re-read, and latency that would pass unnoticed in chat is dead silence on a call. A caller who is misunderstood cannot rephrase in a text box — they repeat themselves more slowly, then ask for a person.