-
-
On 13 tours the model can't have read, Qloo + city size found 13 more of every 100 real cities than size alone; level with the model alone.
-
Every poster is a fan map: eight tours routed on the live site, each printed from its own run on that artist's Qloo heatmap.
-
Asheville's day sheet: the room against the usual crowd, the fans' River Arts District mapped, Kelsey Waldon in support, the hold request.
-
The biggest cities against where Sierra Ferrell's fans over-index on Qloo.
-
Qloo finds, the agent decides, check_plan objects and it re-books: 81 to 104 Qloo requests, 5 or 6 model calls a tour, all in its tour book.
-
The same audience across the rest of Qloo's graph: podcasts, brands, the words fans use, films, TV and books, the sound.
-
Every tour runs twice: 5 of Routed's 8 cities are ones the same model without Qloo didn't pick. Both are scored on Qloo's own fan map.
-
Twenty artists on Qloo's own fan map: Routed against the same model alone, and where the model keeps up.
-
On 49 older real tours, adding Qloo to a big-city list found 7 more of every 100 cities played: 36 artists better, 7 worse.
-
How Routed is built: Qloo supplies the evidence, Nemotron makes the choices, rules check them and send objections back.
-
Presenting a tour: 7 of Sierra Ferrell's 8 stops changed against the same rules on city size alone, with no Qloo data.
Try it: Plan a tour, Sierra Ferrell's tour book, Proof, MCP, GitHub
In 30 seconds
- What it does: from Qloo's taste graph, Routed finds the cities an artist's fans over-index in, the neighbourhood, and the room that fits the crowd, then routes the dates and drafts a checked hold request per venue.
- What we measured: On 13 tours the model can't have read, the ranking Routed ships is level with the model alone (+3.6 pts, not significant) and finds cities the model misses: Asheville is #1 for Sierra Ferrell's fans and #238 by size. Against city size, it found 13.2 pts more (11 better, none worse); Qloo's ranking on its own is 4.6 pts behind the model (not significant).
- A reality check: the flagship is sized for a 1,500 crowd; her 2026 rooms run 1,850–9,525 and 2 of its 8 cities (Portland, the Denver area) are on her 2026 headline dates. Asheville, Qloo's #1 at #238 by size, was her 2025 tour finale; 6 of the 8 are cities she played in 2024–26 (Nashville as support only).
- A design choice: rooms are measured against the crowd, not trusted to any ranking (the model names rooms better, and we publish that).
Inspiration
Independent artists route tours on instinct and a plays dashboard. A language model can recite past tours, not where fans over-index today. Qloo can, so we built a booking agent on it.
What it does
Give an artist, a start city, a date, the shows and the usual crowd. Routed:
- asks Qloo where the fans over-index: one heatmap per region (about 6,000 tiles for North America);
- runs an agent (NVIDIA Nemotron 3 Ultra on Nebius Token Factory) that shortlists cities, scouts rooms and support acts through Qloo, and books them;
- runs check_plan: rules object (radius clause, legs over 1,800 km, room sizes), naming swaps that would clear them, and the agent re-books, up to three rounds;
- measures every room (Wikidata, else a cited web page); a room "fits" only at 85–130% of the crowd;
- maps each city at street level, names the fans' neighbourhood and puts the poster run there;
- writes the hold request a talent buyer wants, every figure checked against the stop's evidence; "Download routing grid (CSV)" and "Copy all holds" take the route to a sheet and every hold into one email.
On the flagship tour, 7 of 8 stops changed against the same rules on city size alone, and 5 of the 8 cities the same model without Qloo didn't pick; it sent her to Toronto, #202 on her fan map. Routed is also an MCP server (fan_map, fan_neighbourhoods, fan_profile, route_tour, get_tour).

What makes it Qloo-powered
Every city, neighbourhood, room, support act and poster run starts from Qloo. The poster wall's 8 tours made 81 to 104 Qloo requests each (median 87) and 5 or 6 model calls, every request listed in its tour book: /v2/insights heatmaps (a WKT polygon per region; POINT(city) and a radius for 150 m street tiles), urn:entity:place by size tier for rooms, urn:entity:artist with signal.location for support acts, and /v2/analysis/compare.
How we built it
Next.js 15 on Vercel, keyless Google Cloud Storage, one Qloo request gate with daily and hourly caps (at the cap, a stored tour); strict-JSON Nemotron 3 Ultra, deterministic rules for what must be exact. 231 unit tests, CI on every push. The public API carries no Qloo ids, raw scores or street tiles.

Challenges we ran into
- Heatmap percentiles are relative to each response: asking for the US and Canada separately floated Calgary to #1 for Tyler Childers. One polygon per region fixed it.
- Qloo tags a 150-cap bar and a 19,000-cap arena alike, so rooms are asked for by size tier, then measured.
Accomplishments that we're proud of
Qloo measured, not asserted: every test published whatever it said, most with the protocol committed first.

How we checked
The first seven rows score cities and rooms against real tours; the street-level pilot uses Qloo's heatmap, the agent-loop row counts check_plan's objections and the last row uses Qloo's own fan map. Artist by artist on /proof.
| Test | Artists | Result | Better–worse–level | Protocol committed before? |
|---|---|---|---|---|
| 13 tours announced after the model's release (unread): Qloo + population vs the model alone | 13 | +3.6 pts (54.5% vs 50.9%), n.s. | 7–2–4 | Yes |
| Same tours: Qloo + population vs population; Qloo alone vs the model alone | 13 | +13.2 pts, p < 0.001; −4.6 pts, n.s. | 11–0–2; 4–8–1 | Yes |
| Same tours: the model + Qloo vs the model alone | 13 | +5.8 pts, n.s. | 7–1–5 | No (post hoc) |
| Unread tours, 25 biggest cities left out: Qloo + population vs population; Qloo alone vs the model | 13 | +12.1 pts; −13.6 pts, n.s. | 8–1–4; 3–8–2 | Yes, from offline caches |
| Club and theatre acts (rooms 3,000 or fewer): Qloo + population vs population; vs the model alone | 20 | +7.9 pts, p = 0.004; −1.8 pts, n.s. | 14–2–4; 5–9–6 | Yes, from offline caches |
| 49 older tours: Qloo + population vs population; vs the model; Qloo alone vs the model | 49 | +7.4 pts, p < 0.001; −3.9 pts; −12.7 pts | 36–7–6; 13–30–6; 7–36–6 | Yes |
| Rooms and openers: Routed's room vs the model's first guess; Qloo's top 3 vs the model's | 34; 22 | real venue in 46 vs 55 of 102 cities (Qloo's first room: 16); top 3: 29 vs 75; opener: 1 vs 6 of 22 | 5–12–17; 1–26–7; 0–5–17 | Yes |
| Street level (pilot): real room vs the city's other rooms, nearer the hottest fan tile | 9 rooms, 3 artists | nearer in 5 of 9, p = 1.00: too few to test; on average 2.3 km farther, and the city centre is nearer in 7 of 9 | 5–4–0 | Yes, from offline caches |
| The agent loop: objections check_plan raised that the agent cleared itself | 9 objections, 4 re-runs | Oct 25: 13 of 15 open. Oct 30 loops: the agent cleared 6 of 9 itself across 4 re-runs (leg cap 4 of 5, radius clause 1 of 2, room 1, major market 1; one new fan-city objection). The one run on the final code (the flagship artist, /tour/d53j244bwp) cleared 0 of 2, settled by rule: it scouted Lexington and Knoxville, booked neither; Louisville stayed inside Cincinnati's 150 km radius, Portland 1,895 km from Denver | 6 cleared, 3 left to the rules, 1 new | No (measured as run) |
| On Qloo's own fan map: Routed vs the model alone | 20 | mean fan rank #8.3 vs #26.2 | No |
The model has likely read the older tours. 13 of the 20 club and theatre room sizes are stated crowds, not measured.
Known limits
- A past tour is one booker's answer; 13 post-cutoff tours is a small sample.
- No avails, guarantees, ticket prices or festival weeks; 1 of 64 rooms on the poster wall has no confirmed capacity.
- Affinity says where fans over-index, not how many tickets sell, and no working booker has validated Routed yet.
What we learned
Adding Qloo to a big-city list finds more of the cities artists really play, on every set we tried; on unread tours it is level with the model and reaches cities the model leaves out. Qloo maps down to 150 m tiles; whether that picks better rooms is untested (9-room pilot).

Who it's for, and what's next
Managers and small agencies who route without a data team (about $29 a seat a month), and promoters. No paying users yet. Next: five working managers using it, their edits published.
A manager's Monday
- Before: the biggest cities and a plays dashboard.
- After: Asheville at #1, and the room measured against the crowd.
- Then: the hold request, pasted into an email.
Built With
- d3-geo
- google-cloud
- mcp
- nebius
- nebius-token-factory
- nemotron-3-ultra
- next.js
- node.js
- nvidia
- nvidia-nemotron
- playwright
- qloo
- qloo-insights-api
- react
- tavily
- typescript
- vercel
- wikidata
