← AEGIS / SHOWCASE

Command Center.

Explore human-authorized coordination in a synthetic incident. You set priorities. The browser suggests routes. Nothing moves without your approval.

Interactive Simulation — No Real Emergency Dispatch

Camera-supported. AI-assisted. Human-authorized.

AEGIS connects computer vision, human assessments, mathematical optimization, neural intelligence, and physical responder hardware to support simulated incident command.

The full HackPrinceton prototype operates locally with actual cameras, backend optimization, trained models, and connected Arduino devices. This public sandbox is an isolated, browser-only representation of that workflow—not the live system.

Begin casualty review ↓
How the Full AEGIS System Works / 8 steps
  1. 01 →Camera observations

    Track people and observable signs; occlusion limits what is known.

  2. 02 →Human confirmation & assessment

    An operator confirms casualties and enters priorities.

  3. 03 →OR-Tools optimization

    CP-SAT produces the authoritative constraint-based proposal; A* plans routes.

  4. 04 →MLP/GNN advisory comparison

    Optional trained models offer comparisons, never dispatch authority.

  5. 05 →Human approval

    The operator reviews and authorizes a specific current proposal.

  6. 06 →Physical Arduino dispatch

    Approved assignments reach connected responder devices.

  7. 07 →ACK / DONE

    Responder acknowledgment and completion events update coordination state.

  8. 08 →Dynamic route updates

    Changed conditions invalidate routes and require renewed authorization.

These stages describe the full prototype. Here, independent browser assignment enumeration and grid shortest paths stand in for backend computation; virtual controls stand in for device events.

Incident

GRID / 04

Confirmed

0 / 4

Active tasks

00

Tasks closed

00

01 / Incident map

7 × 5 / GRID STEPS, NOT REAL DISTANCE
Synthetic incident grid — no real locations or devicesCENTRAL · OPENEAST · OPENC1C2C3C4ABWEST STAGING → EAST ACCESS / SYNTHETIC GRID
● Alpha● Bravo□ Synthetic casualtyDashed route = advisory

Central closure forces a detour. East closure disconnects C2 and C4 from western staging. Crossing controls also support Enter and Space on the map.

02 / Virtual responders

The full prototype integrates physical Arduino responder devices: UNO R4 WiFi — Alpha and UNO Q — Bravo. Devices receive approved assignments and transmit acknowledgment/completion events. Here, receipt, ACK, and DONE are browser-only button simulations; no hardware is connected.

Alpha Idle

Arduino UNO R4 WiFi · virtual only

Bravo Idle

Arduino UNO Q · virtual only

Availability affects new recommendations, never cancels existing work. No hardware is connected.

Camera Intelligence

SYNTHETIC FIXTURES

Illustrative person detections, not camera input or live inference. Frozen scenario snapshot at T+00:12; times below are last-observed offsets, not a live freshness feed. No confidence scores are asserted.

SIM-P01 → C1

T+00:12

SIM-CAM / WEST

Visible

Posture: Lying · Movement: Still

Awaiting explicit human confirmation

SIM-P02 → C2

T+00:10

SIM-CAM / EAST

Partially occluded

Posture: Not testable · Movement: Not testable

Awaiting explicit human confirmation

SIM-P03 → C3

T+00:11

SIM-CAM / WEST

Visible

Posture: Standing · Movement: Moving

Awaiting explicit human confirmation

SIM-P04 → C4

T+00:04

SIM-CAM / EAST

Lost from view

Posture: Not testable · Movement: Not testable

Awaiting explicit human confirmation

A detected person is not automatically a confirmed casualty. Occluded/lost tracks cannot support current sign assessments. Use Human review below to confirm and set priorities.

Recorded CV evidence / separate research clip

Verified recorded output, not this scenario: track t1 in pexels-three-lying-overhead.mp4 first reports lying at 0s, moving at 1s, and still at 2s. These are first-observed clip offsets, not confidence or accuracy scores.

Recorded analysis run: . It is not mapped to C1–C4, used for priorities, or replayed as live inference.

Repository provenance: cv/eval/signs/pexels-three-lying-overhead-20261114T065011.json. Exact source hash is documented in showcase/PROVENANCE.md.

03 / Human review

Confirm each synthetic casualty and choose an operator-defined ordering. These are not clinical triage categories; no diagnosis or treatment is simulated.

C1Courtyard

Synthetic person marker near the central crossing; no clinical assessment.

C2East platform

Synthetic person marker beyond the east access gates; access must be checked.

C3South passage

Synthetic person marker beside the southern alternate corridor.

C4East assembly

Synthetic person marker in the eastern assembly area; awaiting human confirmation.

04 / Recommendation desk

REV 1

Browser algorithm: exhaustive distinct assignment enumeration with breadth-first shortest paths. Maximizes First, then Next, then Later assignments; minimizes total grid steps to break ties. Not live OR-Tools, ML, or a real-world optimizer.

Neural Intelligence / full prototype vs this browser
OR-Tools CP-SAT + A*
Authoritative constraint-based assignment optimization and operational route planning. Only the mathematical proposal can be approved for dispatch.
Pairwise MLP
Learns responder–casualty compatibility scores for constrained advisory matching.
Bipartite GNN
Learns assignments from the graph of responders and casualties, with constrained decoding and independent feasibility checks.

When enabled, the actual backend compares learned advisory recommendations against mathematical optimization. Neural recommendations never authorize dispatch.

This sandbox runs independent assignment enumeration and breadth-first grid routing only. No neural weights, inference, recorded neural scores, or live OR-Tools execution.

No current recommendation. Any confirmation, priority, availability, map, or task change invalidates earlier approval.

Confirm synthetic casualties and set operator priorities to begin.

05 / Virtual task ledger

DISPATCH → RECEIPT → ACK → DONE

Each step is manually simulated. DONE closes a coordination task, not a medical outcome. Blocked tasks cannot progress; reopening a corridor never resumes them automatically.

Awaiting your first authorized dispatch.

Research Intelligence / NYC EMS

RECORDED EVALUATION · NOT EXECUTED HERE

Separate XGBoost research predicts historical incident creation to first assignment duration—not arrival time, route travel, or responder-assignment optimization.

Source incidents
377,606
Held-out incidents
67,798
XGBoost MAE
78.39 seconds
Temporal baseline MAE
82.63 seconds

Aggregate gains were concentrated in 158 rare severity-8 incidents; typical groups showed small gains or regressions. This is not evidence of broad clinical value. The sandbox does not execute XGBoost.

Evaluation scope & provenance

The source snapshot covers January–March 2023, not the entire NYC EMS dataset. After quality filtering and label-availability exclusions, chronological training uses January–February, validation March 1–15, and held-out testing March 16–31. The temporal baseline is the training-only hour/weekday median selected on validation error.

Reported mean absolute errors are aggregate historical evaluation results, not live predictions or a claim of operational readiness. Source: ml/ems/checkpoint/report.json; caveats: ml/ems/checkpoint/generalization_audit.json. Verified excerpts and source hashes are documented in showcase/PROVENANCE.md.