Web3 in Travel

The first build-in-public episode of Web3 in Travel. Instead of telling you what I've learned about AI, I build an AI travel agent on screen, unscripted, with Claude Code and Codex, and you watch every step, including the parts that break.

There are two agents. The first works for the traveller: it gets to know you, then finds where to stay. The second will work for the hotel: it reads the emails of guests who want to book direct, and later it could run on a local AI so guest data never leaves the hotel. The code will be open source, and the agents write the blog about it.

In this episode I test the traveller agent with Lollo, a made-up business traveller, search for three nights in Munich, and check what each search costs. Then I compare the results with Booking.com by hand, and hit the real blocker: the data. The AI was picking the best from a bad list. We fix it live with filters (Google 4.5+, at most €220 a night) and ranking by distance, and find hotels as good as Booking's, one of them €46 a night cheaper.

Collaborations: info@web3intravel.com

CHAPTERS
0:00 From telling to building
7:43 Where we left off
11:51 My setup
22:36 Meet Lollo
29:11 Built in a meeting
30:54 Searching for hotels
37:28 What a search costs
53:18 Booking.com by hand
62:43 The data problem
66:58 The fix

We are in the middle of a bug apocalypse. With AI, finding serious software bugs has become fast and cheap, for attackers and defenders alike. In October a critical PHP bug made governments, ports and hospitals take their own websites offline while they patched.I'm not a developer, yet I found a bug in the Sparrow Bitcoin wallet that the experts had missed. What that says about who builds now, how to attack your own code with AI before someone else does, how I cut my blog's attack surface by moving from WordPress to plain HTML, and why DeFi, the best honeypot for hackers, may come out of this safer than closed legacy software.Collaborations: info@web3intravel.comCHAPTERS0:00 Intro0:20 The bug apocalypse0:51 Both sides got new weapons1:55 The PHP bug3:13 What it means for us3:34 The bottleneck moved5:11 My bug in a Bitcoin wallet6:06 From how to what8:26 Your software is in the war9:24 Attack your own code with AI10:38 Software without bugs is impossible11:35 Fewer doors: the attack surface12:45 My blog: from WordPress to pure HTML15:55 One door left17:22 When non-developers send fixes18:54 Crypto: the best honeypot20:33 Closed software isn't safe either20:54 Immutable contracts21:19 Web3 in Travel's end game?21:39 What to do now

Can a laptop do your company's AI work without sending a single byte outside? Luca De Giglio ran real travel-business tasks on a MacBook Pro with 48 GB of RAM, using the open-weight Qwen 3.8 27B model, and compared every answer with Claude Opus 5.5 through the API.On everyday tasks (reading a booking email, replying to a bad review, pricing a tour package) the local model matched Claude, at about a fifth of the speed and a fraction of the cost, and the data never left the laptop. On a 30-page supplier contract it still found the right cancellation charge, but in six minutes instead of sixteen seconds. On 600 bookings checked against a hotel invoice, it hit a wall.Then the questions that matter for a company: why more RAM is not enough, what actually fits on a laptop, why buying a machine is a bet, who maintains it, and where local AI makes sense in travel and hospitality.IN THIS EPISODE• Three light tests and two heavy tests: local vs. the Claude API.• Why a 64 GB server can be useless for AI, and a 48 GB Mac isn't.• The model as a book: capacity and bandwidth.• Ollama, LM Studio, and models tuned for one machine (DwarfStar, Sushi).• The money: 50× cheaper on electricity, and why that's rarely the point.• Maintenance, evals, and why private isn't automatically safe.• Guest data, contracts and margins: the travel use cases.CHAPTERS0:00 Intro0:20 Local AI: is it time to start?1:43 My laptop: 48 GB of RAM3:16 Why a big server is useless for AI4:08 Test 1: a booking email5:43 Test 2: reply to a bad review7:54 Test 3: a tour package margin9:22 Speed: tokens per second10:12 Cost, and where the data went11:25 Heavy test A: a 30-page contract14:05 Heavy test B: 600 bookings vs. an invoice16:43 Electricity: 50× cheaper17:25 The model is a book19:01 What runs on 48 GB, and the tools19:56 Models tuned for one machine22:29 Does local save you money?23:57 Buying a machine is a bet26:09 Finding the ceiling, and evals28:07 Maintenance28:53 Private isn't automatically safe29:32 Local AI in travel and hospitality30:57 A possible setup, and where to startFor collaborations, advisory work or speaking:info@web3intravel.comhttps://www.linkedin.com/in/luca-de-giglio/

Most companies are still asking how to use AI. The next question is how to use it privately. When your team plugs Claude or ChatGPT into the company's documents, where does that data go, and what can the AI labs do with it?In this episode of Web3 in Travel, Luca De Giglio explains why your data is worth far more to an AI lab than it ever was to Google, why a $20 subscription buys hundreds of dollars of inference, and what "we don't train on your data" really promises. He then walks through the options for private AI, from the most private to the least: local open-weight models on your own hardware, private APIs for open models, and Trusted Execution Environments (TEEs). For each, what it costs and what it actually protects.He closes with the strategy: route each task to the right model, decide which data must never leave the company, and accept that total privacy does not exist. You manage risk, you don't cancel it.IN THIS EPISODE• Why LLMs extract more value from your data than ads ever did.• The real economics of a $20 AI subscription.• Could an AI lab become a competitor in your industry?• Local models and open weights: DeepSeek, Kimi K3, Qwen, gpt-oss.• Private APIs and anonymized inference: what they cost, how private they are.• TEEs: hardware attestation, and why you still have to verify.• Routing: the right task to the right model.• The first step: decide which data never goes near an LLM.MENTIONED• Vitalik Buterin on a local model deciding what to send to frontier models: https://x.com/VitalikButerin/status/2106537633056969024CHAPTERS0:00 Intro0:20 The next question: how do I use AI privately?0:59 Google sold ads. LLMs take the value1:33 Your best insight, in the next model2:39 Why a $20 subscription is so cheap4:34 Do they train on your data?5:24 When the AI lab enters your industry7:51 Local models and open weights10:07 The cost of running AI yourself11:25 Private APIs: paying full price13:08 How private is anonymized?14:21 TEEs: trusted execution environments16:12 Routing: the right task to the right model17:43 A strategy for your data19:48 The first step: data that never leaves21:21 Nothing is 100% privateFor collaborations, advisory work or speaking:info@web3intravel.comhttps://www.linkedin.com/in/luca-de-giglio/

The real leverage from AI starts when you go beyond ChatGPT: a project with agents for each area of your work, in tools like Claude Code or Codex that build their own tools.

I share three people I onboarded and what changed for them, why these coding tools are for non-coders too, and why the terminal stops most people. I introduce The White, the software I use to onboard people, and explain who succeeds, who fails, and why you should start small.

Want help making the jump?
info@web3intravel.com
https://www.linkedin.com/in/luca-de-giglio

CHAPTERS
0:00 Intro
0:20 Why I share new things early
1:35 Beyond ChatGPT: a project for everything
3:35 How do you make the jump?
4:29 Three people I onboarded
5:30 Coding tools are not just for coders
6:38 A third way to get software
7:55 How one CEO automated his company
10:45 "What did you do to him?"
13:57 The terminal problem and The White
16:18 An offer for listeners
17:24 Who succeeds
19:20 Why AI companies want you out of the chat
21:23 Who fails
23:03 Start small
24:29 Not a CEO? Become a secret superhero