Fifty thousand faces. One camera. One night.
Pavilion deployment — client name pending publication
Hero visual placeholder · CamBot
Challenge
A twelve-night activation inside a pavilion at a Riyadh event. Footfall projection: fifty thousand across the run.
The brief was a single sentence. Every visitor walks away with a cinematic AI-generated piece of content featuring them — under thirty seconds from approach to shareable file. No queue longer than two minutes at any point of any night. The brand wanted personalized output that did not look algorithmic.
The team had four weeks.
System
We built CamBot.
A self-contained AI camera installation that recognizes a visitor, captures three frames in two seconds, runs an on-device generative pipeline to produce a six-second cinematic in the brand's visual language, and delivers it to the visitor's phone via QR before they have walked five steps from the column.
The visitor side is a single physical surface — a backlit camera column with the brand mark, a soft activation zone for stepping into, a delivery screen showing the generated frame.
The studio side is what nobody sees. An edge inference rig running our generative model fine-tuned on the brand's visual reference set. A real-time face-aware crop and stabilize pipeline. An operations dashboard showing per-station throughput, queue depth, model latency, and content moderation flags. A cloud handoff that names, brands, and ships the output to the visitor's phone via a one-tap QR code.
The whole system runs offline-capable. Latency target: twenty-two seconds end-to-end. Achieved: nineteen.
Outcome
Across twelve nights, CamBot processed forty-seven thousand visitors and delivered forty-four thousand personalized cinematics. Average throughput per station: sixty-two visitors per hour. Average time to delivery: nineteen seconds. Peak queue: six people.
The brand received the activation's content directly from the audience's phones — eleven thousand organic shares across Instagram and TikTok within seventy-two hours of each visit. No downtime across the run.
The activation became the most-photographed point in the pavilion. The studio became the layer underneath.
By the numbers
47,300
visitors processed
94.7%
delivery rate
19s
average end-to-end latency
11,200
organic social shares within 72h
0
minutes of downtime
6
maximum queue length at peak
Numbers shown are illustrative for the sample showcase. Real values will be supplied for publication.
Tech stack
Hardware
Backlit camera column (custom fabrication), NVIDIA Jetson Orin edge units, professional capture rig.
Vision pipeline
Face detection and stabilization (custom on PyTorch).
Generative model
Brand-fine-tuned diffusion (Flux base with custom LoRA trained on the brand's visual reference set).
Real-time orchestration
Trigger.dev workflows + Cloudflare Workers for edge-to-cloud handoff.
Delivery
Cloudflare R2 for storage + Stream for adaptive video, QR-issued short links via Cloudflare Workers.
Operations dashboard
Next.js 16, HeroUI, Supabase Realtime.
Moderation
In-house safety classifier with a human review queue for edge cases.
Timeline
Brand model training: two weeks. Hardware deployment: one week. On-site setup: three days.
“We did not know it could be this fast and this beautiful at the same time.”
Pending real attribution — quote available on request.
Team and credits
Humans
- Creative direction[Name]
- Hardware integration[Partner]
- Engineering lead[Name]
- On-site production lead[Name]
- Brand model training[Name]
- Photography & film[Name & studio]
Agents
- Sirius ConductorLed the project end-to-end.
- Sirius BuilderBuilt the vision pipeline, edge inference rig, and operations dashboard.
- Sirius ComposerSupervised the brand-fine-tune and concept directions.
- Sirius VoiceComposed the showcase narrative.