Ten thousand people. One venue. Twenty minutes ahead of the bottleneck.
Concept demonstration — built against synthetic crowd simulations and an anonymized venue plan.
Hero visual placeholder · CrowdFlow
Challenge
Most venues find out about a bottleneck the way the crowd does — when it is already happening.
By then the staff has already failed to redeploy, the queue has already formed, the social posts about the wait are already out, and the dashboards in operations are still showing footfall from twenty minutes ago.
We wanted to test whether a small, fast system could see what was coming before the venue did.
System
We built CrowdFlow.
A predictive layer that ingests anonymized vision feeds and edge-sensor density from across a venue, fuses them with the day's schedule and historical patterns, and outputs three things the operations commander actually needs: where the crowd is now, where it will be in fifteen minutes, and what to do about it.
The visitor side is invisible by design. No app. No QR. No data tied to a person. The system reasons in densities, not identities.
The operations side is one screen. A live heat plan of the venue. A timeline showing predicted congestion thirty minutes out. A queue of suggested interventions — open this gate, redeploy that team, route signage through that corridor — each one a single click to accept.
Inference runs on edge boxes co-located with sensor clusters. The model is a small ensemble — a convolutional density estimator plus a temporal forecaster trained on synthetic crowd dynamics and a corpus of public venue footage. The prediction loop closes in under thirty seconds.
Outcome
Across the simulated run, CrowdFlow predicted ninety-three percent of bottlenecks at a fifteen-minute horizon and seventy-eight percent at a thirty-minute horizon. The operations dashboard surfaced an actionable suggestion an average of seventeen minutes before the bottleneck would have formed.
The concept demonstrates the operating shape Sirius proposes for any venue with more than a few thousand visitors at peak: a thin predictive layer, calibrated to the specific venue and event mix, that lives entirely on the operations side and never touches the visitor.
We are now in conversation with two venue operators about a real deployment.
By the numbers
Modeled: 93%
bottleneck prediction at 15min
Modeled: 78%
bottleneck prediction at 30min
Target: 17min
median actionable lead time
Target: <30s
end-to-end prediction loop
0
personal data captured
Concept
real deployment pending
Concept demonstration. Modeled and target values shown above were produced against synthetic crowd simulations and a public venue footage corpus, not a live deployment.
Tech stack
Sensing
Anonymized vision feeds (density heatmaps only) plus edge-sensor occupancy counts. No identification, no tracking, no PII.
Density estimation
Custom convolutional density head trained on synthetic crowds with sim-to-real domain adaptation.
Forecaster
Spatio-temporal transformer with venue-aware positional encoding and event-schedule cross-attention.
Edge runtime
NVIDIA Jetson Orin units co-located with sensor clusters. Inference latency under two hundred milliseconds.
Operations dashboard
Next.js 16 + Supabase Realtime + Mapbox custom-styled venue overlays.
Training data
Synthetic Unity crowd simulations + a curated corpus of public venue camera footage with explicit licensing.
Timeline
Two weeks for the concept build, end to end. A real deployment scopes at six to ten weeks per venue.
Team and credits
Humans
- Concept direction[Founder]
- Engineering lead[Name]
- Simulation design[Name]
Agents
- Sirius ComposerDesigned the operations dashboard and the venue heat plan visual language.
- Sirius BuilderTrained the density head and the forecaster, built the edge runtime.
- Sirius ObservatorySourced public footage and built the curated training corpus.
- Sirius VoiceComposed the showcase narrative.