Above China, Your Starlink Doesn't Exist

From the Travel Tech Podcast newsletter for episode 31, Above China, Your Starlink Doesn't Exist.

Somewhere above Chinese airspace, a passenger asks for a glass of water and the answer, technically, is no. Not because the crew is out of water, but because the point-of-sale system has no connection, no matter how much bandwidth Starlink promises elsewhere. I put that story in front of the readers of this newsletter because it's the clearest example I've heard of a problem that looks technical and is actually geopolitical. I wanted that perspective on the show.

He started by building point-of-sale software for a 2,500-outlet coffee chain in India, on Android, before most retailers had reliable internet. That offline-first habit turned out to be exactly what in-flight commerce needed, and it pulled him into aviation almost by accident, into patents on offline card fraud detection, and eventually into an AI agent that gives ground staff cited, evidence-backed answers instead of a 2,000-page SOP and a guess. Apple Podcasts, also streaming on Spotify and YouTube.

The tension running through the conversation is this: every airline wants AI that staff and passengers can trust, but trust requires evidence, and most AI tools are built to sound confident rather than to show their work.

That Tension Has a Structural Cause

That tension has a structural cause. In-flight systems can't assume a network connection exists, because access to satellite and cellular data is restricted by individual governments regardless of the technology available, which means every transaction, every SOP lookup, and every AI query has to work offline first and reconcile later, across potentially thousands of aircraft syncing data after landing.

AI adds a new layer of pressure on top of that constraint. Ground staff handling multiple airlines can't memorize thousands of pages of airline-specific compliance rules, and Ajith's team has found that citing the exact SOP page and paragraph behind an AI answer is what actually earns staff trust, more than the answer itself. Forthcode has avoided routing that data through frontier LLMs entirely, choosing to host smaller models locally so airline data never leaves the environment, and their next release in September is a fully offline AI system built to run inside the aircraft itself.

Ajith Balakumar, CEO and founder of Forthcode, joins me to talk through the accidental path from coffee shop POS systems into aviation, the offline-first architecture behind Forthcode's retail, catering, and Lounge in the Sky products, the AI maturity curve from co-pilots to operational intelligence, and how he uses AI to build fully working prototypes in about an hour as a founder.

Key Takeaways

• Evidence beats confidence: AI earns trust from frontline staff when it cites the specific SOP text behind an answer, not when it simply sounds certain.
• Offline-first is a business model, not just an engineering choice: Forthcode's coffee-chain POS background became the foundation for every aviation product it built afterward.
• AI maturity has three stages: co-pilots where humans decide everything, agents where humans approve a subset of decisions, and operational intelligence where multi-agent systems inform executive decisions.
• Regulation is heading toward plug-and-play compliance: expect cloud providers to offer standardized AI guardrail layers, similar to an API gateway, that route traffic through GDPR- and SOC 2-compliant models automatically.
• Founders can now out-build their own specs: working prototypes that used to take days of back-and-forth can be built in about an hour, changing how founders communicate vision to their teams.

🔧 Two to Build

This week's episode makes the case that ground staff are often the last line of defence when connectivity, tools, and managers are all unavailable at once.

Build a cited SOP assistant that ground crew can query in plain language, without a manager in the loop.

Most airline SOPs live in PDFs that nobody searches well under pressure. An AI assistant trained on those documents can return answers with the exact page and paragraph, so staff trust the output and auditors can verify it. Forthcode built this as their Davix agent, running across multiple airlines simultaneously. The citation layer is what separates a useful tool from a liability.

Catch: the hard part is not the retrieval. It is keeping the document corpus current when airlines update procedures partway through the season and nobody tells your ingestion pipeline.

→ AI ground crew assistant

Build a decision support layer for irregular operations that hands ground staff a complete entitlement package before they ask for one.

When a flight cancels, the clock starts immediately and passengers expect answers in minutes. A system that reads the disruption event, checks the passenger booking class and fare rules, and surfaces the correct meal voucher, hotel tier, and rebooking options removes the need to escalate. Davix does exactly this, which means a junior agent can handle what previously required a duty manager. The cost saving is real, but the bigger gain is consistency across a shift.

Catch: entitlement rules vary by route, fare family, and regulation, so a model that generalises across all three without fine grained rule tables will confidently give passengers the wrong hotel.

→ IROP decision support

Fly safe,
Alex Brooker, Host of the Travel Tech Podcast, Founder of Airside Labs