How Kindling works

The problem, in numbers

• 6%

  of restaurant reservations are for Monday nights, and 8% for Tuesday (Toast data).

  Source: https://www.chowhound.com/1841628/why-restaurants-closed-on-mondays

• 1 in 6

  people is affected by loneliness (WHO Commission on Social Connection, 2025).

  Source: https://www.who.int/groups/commission-on-social-connection

• 50 h

  of time together turns an acquaintance into a casual friend (Jeffrey Hall, 2018). One-off events cannot reach that. A weekly club can.

  Source: https://news.ku.edu/2018/03/06/study-reveals-number-hours-it-takes-make-friend

The weekly loop

1. 1

   Scout

   Reads your city's taste, the places within 1 km and your venue's crowd with Qloo.

2. 2

   Design

   Proposes hidden clubs from that evidence.

3. 3

   Verify

   Checks and scores each club with more Qloo calls.

4. 4

   Produce

   Writes the kit: poster, run of show, listing, menu.

5. 5

   Seat

   Seats guests by taste at check-in.

6. 6

   Coach

   Reads the week and says keep, tweak or switch.

Then it goes round again. The same people come back, the Coach reads who came and who returned, and next week's plan changes because of it. That repeat is the point: friendships need hours together, and one night cannot give them.

The agent roles

Locator

Finds the venue on Qloo and reads its own tags.

Scout

A model-driven tool loop. It chooses which Qloo scans to run, widens the radius when data is thin, and records a reason for each call.

Designer

Proposes 6 to 8 clubs from the Scout's signal pool. Every club must cite things that are really in the pool, and anything else is dropped by code.

Verifier

Checks the top clubs with a coherence call and scores all of them with a fixed formula.

Producer

Writes the club kit. The playlist is picked by code from the pool, and a licence note is added whenever a text mentions screening a film.

Seater

Pure code. Weighted tag overlap decides who sits together, and nobody is left alone.

Coach

Fixed rules decide keep, tweak or switch from the real numbers. The model only writes the wording.

Baseline

Asks a plain language model the same question with no Qloo data, so you can compare.

Each stage is its own request and writes a plain-language trace, which is what you see on the run page.

What is Qloo, what is a language model, what is our code

Qloo

What the area loves, what is rising, which clubs fit together, each guest's taste vector, nearby places and the heat map.

The language model

Wording and creative choices: club titles, the kit text, the Coach's variant, and the comparison with no Qloo data. It never sees a guest's name or picks.

Our code

The club score, the seating, the keep, tweak or switch decision, the regulars count, the licence note and every check that a cited thing exists.

Every Qloo call, and why

All calls go through one client that paces them, retries once, caches them and logs each one. The counts below come from that log (459 calls so far), not from a script. The picks page shows the exact calls behind each score.

How far each signal reaches

Qloo reads film, music, TV, book and podcast taste for a whole city, so that part is the same for every venue in it. What makes a club fit your venue is the other two.

• Your venue's crowd

  The tags Qloo holds for your venue and for venues like it. Different venue types have different crowds, so a brewery and a cafe get different clubs.

• Places within 1 km

  The places around your door, with their categories. This is the most local thing Qloo gives, and it differs from one neighbourhood to the next.

• Your city's culture

  What your city loves and what is rising there. It tells us which films, songs and books are worth building a club on.

The club score

A fixed formula, not a model's opinion. The weights are below, and each part comes with a reason and the Qloo calls behind it.