About Silicon Islands AB
A long-form account of why the hardest problem in hospitality was never the data, what PlaceProfile built, and how the pipeline runs end to end.
Part one
There is a concept in theoretical computer science that most product teams will never encounter in a textbook but live with every day in practice. It is the gap between finding and verifying. Checking whether a hotel recommendation is good is easy. A guest either felt it or did not. Producing that recommendation in the first place, from thousands of variables, competing signals, and subjective human preferences, is the hard part.
For decades, the industry accepted this asymmetry as a fixed condition and built accordingly. The result was a generation of place-intelligence architectures built around approximation. Star ratings that compress nuance into a single integer. Review aggregates that flatten contradictory experiences into a mean score. Recommendation engines that optimize for engagement rather than fit.
These systems are not failures. They are rational engineering responses to a genuinely hard computational problem. When finding the right answer is expensive, you build a machine that quickly finds a good-enough answer. PlaceProfile took a different starting point.
Part two
PlaceProfile's core insight is that the approximation problem in hospitality was never really a data problem. It was a language problem. The industry had abundant data and impoverished descriptions. It knew how many square meters a terrace occupied, but not whether sitting on it felt expansive or intimate. It tracked check-in times but not whether the arrival experience felt considered.
By focusing on how to describe a place's experiential texture in language that AI systems can interpret and carry, PlaceProfile effectively moved intelligence from the retrieval layer to the description layer. The place profile itself becomes the solver. When an AI assistant processes a guest query, a well-constructed PlaceProfile description does not need to be approximated or inferred. The resonance between the query and the description is direct.
This is not a marginal improvement in matching accuracy. It is a structural change in how the search problem is framed. The old architecture asked AI to infer feelings from attributes. PlaceProfile gives AI the feeling directly, expressed in the language AI was built to work with.
Part three
A single user action — paste a playlist URL and click import — triggers a deterministic pipeline. The playlist resolver fetches the track list. Public Spotify playlists, public Deezer playlists, and the playlist service that ships under the placeprofile.net subdomain are all auto-detected at the URL level. The output of this step is an ordered list of track identifiers paired with track metadata.
Each track is then resolved against the Spotify audio-features layer. The dimensions used are the published audio features Spotify exposes via its public Audio Features API:
In parallel, each track that has an ISRC is matched against the AcousticBrainz mood-classifier project. Six mood dimensions are used, each a 0–1 confidence score:
The fingerprint computer aggregates per-track data into a venue-level signal at multiple slices — venue-wide, per-playlist, and per-daypart. The daypart breakdown matters because most venues do not sound the same across the week. Six dayparts segment the week:
The graph builder assembles a schema.org @graph from the venue's metadata, the venue-wide and per-daypart fingerprints, and the per-track resolution detail. The output is roughly 93 kilobytes for a typical venue with a 100-track playlist. The publishing layer renders the graph as application/ld+json embedded in the head of an HTML page at placeprofile.net/v/{slug}. The page body is server-rendered, so AI bots that fetch the URL receive a complete document on first response, with no JavaScript execution required.
Part four
If a daypart slice has fewer than approximately 15 percent of its tracks with mood data, the mood vector is omitted from the published profile for that slice. The audio-features vector publishes regardless. The result, when coverage is thin, is an audio-features-only fingerprint with an explicit signal in the dashboard that mood is unavailable at that scope and why. Under no circumstances does the system publish a misleading mood average over a tiny sample.
Part five
Bohuslän constitutes the strategically defined entry market. The total number of indexed hospitality venues in the Bohuslän region is 726. These venues, a concentration of independent restaurants, cafés, bars, and visitor destinations along one of Sweden's most visited coastlines, are exactly the customer profile the product is designed for: atmospheric, music-programmed, independent, and competing for visibility in an AI-discovery layer currently dominated by chains.
The assumption for the Bohuslän pilot programme is that 50 percent of the 726 venues subscribe via the tourist bureau in Stenungsund under a partner-priced agreement of SEK 7,000 per year. That produces 363 paying customers and total recurring revenue of SEK 2,541,000 per year from the Bohuslän region alone.
The phase 1 addressable market at national and Nordic level consists of 5,822,075 indexed venues, of which 699,351 qualify for outreach across Sweden, Denmark, Norway, and the United Kingdom. Bohuslän is not the total market. It is the validation anchor that demonstrates the model works before it is rolled out at broader scale.
Part six
The hospitality operators who understand this shift earliest will hold a durable positional advantage. Investing in the linguistic quality of place descriptions is not a marketing exercise. It is infrastructure. As AI assistants become the primary interface through which guests discover, evaluate, and choose properties, the description layer becomes load-bearing in a way it never was before.
PlaceProfile's contribution is not simply a better profile format. It is a recognition that the asymmetry between finding and verifying, the gap that defined a generation of search architecture, collapses when the description itself conveys the feeling. At that point, the AI assistant is not approximating an answer. It is reading one.
The company
Legal name: Silicon Islands AB
Organisation number: 556974-8865
Registered address: Nedre Tådås 13, 47192 Klövedal, Sweden
Contact: [email protected]
Role: Legal controller and operating company of placeprofile.net.