Out-of-home planning by taste · Atlanta sample

Put the billboard where the fans live.

Wallspace picks the billboards, bus shelters and digital screens where a brand's audience actually lives, using Qloo's taste data instead of phone-location tracking, and writes a line of copy for each face from what that block loves. A planner gets a ranked, budgeted plan in about a minute. A media owner gets proof of who lives around each face.

Open the sample Atlanta plan For media owners

Aggregate taste by place: no device tracking, never identity. The brand (Porchlight Comedy Tour), the added faces and all rates are samples, and each is labelled.

The problem, for both buyers

Brands, agencies and touring acts buy out-of-home on traffic counts and phone-location panels. Those say who drove past a board, not whether the people who live around it love what is being sold, so faces get picked on impressions and gut and every board runs the same creative.

Independent media owners with 20 to 500 faces compete with Lamar, Outfront and Clear Channel, who have audience-data teams. They sell on location and price because they cannot show who lives around a face.

How it works

Taste by place. Describe the audience by what it loves. Qloo maps where that taste concentrates across Atlanta and what each block over-indexes on.
The agent picks and writes. A Nemotron agent on Nebius Token Factory scores every face by its block, reads site photos for sightline and clutter, checks local events during the flight, writes three lines per face and splits the budget. Nemotron Ultra reviews the finished plan once.
Into the tools you already use. The plan goes out by email, as a Slack approval, as a CRM deal, as a sheet with one row per face and as flight dates on a calendar. Replies, approvals and edits come back to the plan.

Every run step names the model that ran it, with tokens, time and cost from the run log. See it on the sample plan.

The stack, said plainly

Qloo Taste AI: What it is: A cultural-affinity engine: what people who love one thing also love, aggregated by place, with no personal data.; What it does here: Where the audience lives (heatmap), what each face's block over-indexes on, and which brands' fans over-index around a face.
NVIDIA Nemotron on Nebius Token Factory: What it is: NVIDIA's open Nemotron models, served by Nebius as a pay-per-token API.; What it does here: Nemotron 3.5 Super VL runs the planning agent and reads site photos; Nemotron 3.5 Lightning writes copy; Nemotron 3 Ultra reviews the plan. Token Factory also serves Qwen3-Embedding-8B (not an NVIDIA model), which maps free-text briefs to Qloo tags.
NeMo Guardrails on Nebius Serverless Jobs: What it is: NVIDIA's open-source toolkit for rails around a model, run as a batch job on Nebius AI Cloud.; What it does here: Signs off every copy line before print against a taste-only policy; a red-team mode publishes catch and false-positive rates on the limits page. We start the job from the Nebius CLI, on demand.
Tavily: What it is: A search and extraction API built for AI agents.; What it does here: Finds dated local events near each picked face during the flight. Results are cached per neighbourhood and month.

Hosted on Cloudflare Workers. Maps by MapLibre and OpenFreeMap. Customer tools connect through Composio and a Wallspace Slack app.

For media owners

Media-owner mode turns the inventory around: for each face, what the people who live around it love, and the brands whose fans over-index there, ready to pitch.

Open media-owner mode

What is real and what is sample

Real: billboard positions from OpenStreetMap, City of Atlanta neighbourhoods, model runs and their costs, Tavily events with source links, guard results.
Sample: the brand, faces added where OpenStreetMap has none, face sizes and rates, and example site photos.
Qloo taste results are cached; each evidence chip says whether it is cached or live.

Read the limits page for real numbers on cost, dropped faces, guard misses and coverage gaps.