An AI mobile training platform for SANA Hotels — from PowerPoints and PDFs to gamified mobile courses

SANA Hotels operates across Portugal and internationally, with staff across front desk, housekeeping, food and beverage, maintenance, and management. Each group needs ongoing training on brand standards, operational procedures, safety protocols, and guest experience guidelines.

The model SANA was running was small-team and hands-on. Trainers travelled from hotel to hotel every month delivering sessions in person from PowerPoint decks. When the group was smaller, this worked. As the portfolio expanded, the math stopped working. Every new property added another stop on the circuit. Every new hire triggered another in-person onboarding session. Every policy update required re-travelling to every location. The team was spending more time in transit than developing content. And the workforce was projected to grow 20% by 2027.

There was a quieter consistency problem too. When training depends on a person delivering it live, content drifts. A trainer covering the same material for the fifteenth time in a month adjusts, abbreviates, skips sections. New hires at different properties end up with subtly different versions of the same training. For a hotel group where brand consistency directly affects guest experience, that's a slow-moving but real risk.

And then there was the content itself. SANA's training library — hundreds of PowerPoint decks and Word documents accumulated over years — represented a genuine investment in institutional knowledge. But the format made that knowledge hard to use. Employees couldn't review materials between shifts. There was no way to check comprehension. No way to track completion. Translating across languages required manual effort for every document.

What we learned

Twistag built a platform that converts SANA's existing library into three output formats from a single document upload: text-based interactive courses with quizzes, video lessons presented by AI-generated avatars, and personalised learning paths scoped by role, hotel location, and language. The pipeline runs entirely on serverless AWS infrastructure with AWS Bedrock and Anthropic Claude doing the AI work.

The platform separates cleanly into three layers. Content ingestion and generation takes documents in and produces structured course modules. The learning platform handles how employees access, progress through, and are assessed on those modules. The personalisation layer determines which modules each employee sees based on role, location, and language. Separating these concerns was deliberate — it means the AI generation pipeline can improve over time without changing how employees experience the front end, the personalisation logic can evolve without touching the underlying content, and each layer can be debugged and optimised independently.

Converting a corporate PowerPoint into a structured learning module isn't a single AI call. The pipeline runs in stages. First, the system ingests the document and extracts its structure — headings, bullet points, images, speaker notes, embedded media. This step has to handle the messiness of actual corporate files: inconsistent formatting, slides that are mostly images, mixed-language content within a single deck, information that lives in speaker notes rather than visible content. We wrote platform-specific extraction logic for PowerPoint and Word separately, because the structural conventions of each format are different.

Once raw content is structured, it passes through Bedrock with Claude for a sequence of AI tasks. The model identifies core learning objectives. It restructures content into a logical lesson flow for mobile consumption — shorter sections, clearer progression, logical grouping. It generates scenario-based quiz questions that test application rather than memorisation. And it produces metadata tags for personalisation: which roles, which properties, which language. Running these as sequential steps rather than a single combined prompt produces more reliable output. Each step receives clean, structured input from the previous one.

The same source document can produce three distinct outputs depending on what the content needs. Text-based interactive courses are the default — five-to-ten-minute modules with comprehension questions interspersed using spaced retrieval principles. Video courses use AI-generated avatars to deliver content where tone and delivery matter as much as the words: brand standards, guest communication training, property introductions. The personalisation engine then assembles each employee's learning path from the course library, filtered by their role and property. Same fire safety document, two versions — one emphasising evacuation communication for guest-facing staff, one emphasising equipment shutdown for maintenance — produced from a single upload.

What this shaped

Training that previously required scheduling and a train ticket now reaches every employee at every property within minutes. The training team's time has shifted from logistics to content development. Trainers off the road. Content in pockets. Comprehension tracked. Languages handled.

The platform was designed to absorb the projected 20% workforce growth without adding training staff. Same team, same content engine, more properties and more employees feeding through it. The institutional knowledge in those hundreds of PowerPoint decks is now accessible, assessable, and scalable — without any of it being rebuilt from scratch.

What this proved