AEGIS
Hackathon Build

A shield of perception
for the medic on the ground.

AEGIS helps one combat medic see more, hear more, and prioritize faster during MASCAL events. The system keeps casualty identity stable, estimates injury burden, and highlights who needs urgent care first without taking the human out of triage.

Live scene readout
Focus casualty: A1
Immediate

Visible torso hemorrhage and multiple wound sites push A1 to the top of the attention stack while the remaining casualties stay tracked in the scene.

Tracked
03
Bleeding
02
Doctrine
SALT
Vision

Tracks casualties, scores visible wound burden, and surfaces active hemorrhage first.

Audio

Flags respiratory distress and airway compromise when the scene is too loud to trust the ear alone.

Triage

Builds medic-confirmed SALT/TCCC suggestions and drafts MEDEVAC from the live roster.

Aligned with doctrine
  • SALT Triage
  • TCCC
  • 9-Line MEDEVAC
  • DIL environments
  • SOCOM-grade doctrine
Perception

Every wound seen.
Every casualty tracked.

The vision pipeline runs zero-shot on tactical edge hardware. It detects every casualty in frame, holds identity across chaotic scenes, and measures wounds the moment they come into view — no custom training, no cloud.

Detect & track
YOLOv8 with ByteTrack maintains persistent track IDs across partial occlusion, smoke, and low light.
Segment in situ
MobileSAM and Grounding DINO localize and measure wounds on the fly — arterial bleeds get surfaced first.
Re-identify after loss
DINOv2 embeddings recover the same casualty when they leave the frame and return, so the roster never breaks.
vision · frame 00:04:12
8.2 cmCASUALTY #07 · 0.94wound · 0.87 · hemorrhagere-ID: casualty #7YOLOv8 · ByteTrackMobileSAM · Grounding DINO24.7 fps
Detect24.7 fps
Segment0.87 conf
Re-ID312 frames
Acoustics

Listening for what
the noise hides.

In a MASCAL scene the ear is overwhelmed long before the eye. AEGIS runs zero-shot audio classification continuously and gives the medic their hands back with voice-first command, recognized at the edge.

Zero-shot acoustic triage
CLAP flags stridor, gurgling, agonal breathing, and absent respirations with calibrated confidence scores.
Respiratory rate, continuous
Cycle detection estimates breaths per minute from ambient audio — surfaced alongside the visual track.
Voice command, hands-free
Whisper parses tag calls, interventions, and MEDEVAC requests. Every command round-trips through confirmation.
audio · 00:00:06.4
Stridor detected· 0.87
CLAP zero-shot
00:00.000:03.200:06.4
Resp rate34br/min
PatternIrregular
Pulse oxspO₂
Whisper · transcript
03.1MEDICred tag 7
parsed intentset_triage(id=7, tag=IMMEDIATE)
03.4AEGISIMMEDIATE assigned to casualty 7 · confirm?
Triage & MEDEVAC

Doctrine in the loop.
Medic on the trigger.

The fusion engine reconciles vision, audio, and voice into a single casualty record aligned with SALT and TCCC. Every AI-derived field carries a confidence score. Nothing is committed without the medic's hand.

SALT-aligned suggestions
The engine proposes IMMEDIATE / DELAYED / MINIMAL categories from the fused signals — the medic confirms, amends, or overrides.
Interventions, logged
Tourniquets, airway management, and pulse checks are captured by voice or tap with time-stamped provenance.
9-Line MEDEVAC, pre-filled
Lines 1–8 draft automatically from the roster and scene metadata. The medic reviews the nine lines, not assembles them.
casualty #07 · expanded
IMMEDIATE · SALT

Casualty #07

11:04:22
AI · 94%
MechanismPenetrating · R thigh
AI · 91%
HemorrhageActive · arterial
AI · 94%
AirwayPatent
medic · manual
Respiration28 br/min
AI · 83%
Pulse132 bpm (radial)
medic · manual
Interventions logged
  • 11:04MEDICCAT tourniquet · R thigh proximal
  • 11:05VOICEAirway clear · no occlusion
  • 11:06SYSTEMMEDEVAC request generated
9-Line MEDEVACDRAFT · unconfirmed
  • 01
    Location
    38TKL 042 119
  • 02
    Freq / callsign
    45.50 · DUSTOFF 3
  • 03
    Casualties by precedence
    A×2 B×1
  • 04
    Special equipment
    Ventilator
  • 05
    Casualties by type
    L×2 A×1
  • 06
    Security at PZ
    Possible enemy
  • 07
    Marking method
    awaiting input
  • 08
    Nationality / status
    US military
  • 09
    NBC contamination
    awaiting input
Medic must confirm before send
How it works

Four stages, one loop,
always medic-confirmed.

Detect & track

Every casualty in frame gets a persistent track ID the moment they appear. Smoke, partial occlusion, and re-entry don't break the roster.

Assess & segment

Wounds are localized and measured. Respiratory sounds are classified. Posture, motion, and pulse-present cues are fused per casualty.

Confirm & triage

The fusion engine proposes a SALT category with confidence. The medic confirms by voice or tap. Nothing is committed autonomously.

Evacuate

Once triaged, a 9-Line MEDEVAC drafts itself from the roster. The medic reviews, edits, and sends — mesh, radio, or runner.

Ethics & safety

Perception augmentation,
not autonomous triage.

AEGIS never makes a life-or-death decision. It surfaces information faster so the medic can. These are the guardrails enforced in code, not policy.

Guardrail

Every suggestion requires medic confirmation.

Triage categories, interventions, and MEDEVAC sends all round-trip through an explicit human confirmation step. AEGIS proposes — the medic decides.

Guardrail

Expectant (black) cannot be AI-assigned.

The system is prohibited in code from suggesting an expectant categorization. That determination is medic-only. Always.

Guardrail

Every AI-derived field carries a confidence score.

Wound masks, respiratory classifications, and triage proposals display calibrated confidence. Low-confidence signals surface as prompts, not claims.

Guardrail

Full local audit log.

Every suggestion, confirmation, override, and voice command is written to a local SQLite log with timestamps and provenance — for after-action review.

Guardrail

One-toggle manual override.

AI suggestions can be disabled entirely with a single switch. The dashboard keeps working as a structured casualty tracker without any autonomy.

Guardrail

Fully offline operation.

AEGIS runs on a 20–40 W edge device with no network dependency. No cloud, no uplink, no data leaves the device — by default, not by configuration.

FAQ

Honest answers about
a high-stakes tool.

If something here isn’t covered, open an issue on the repo — we’d rather be direct than handwave.

No. AEGIS is perception augmentation, not autonomous triage. Every suggestion — SALT category, intervention log, MEDEVAC send — requires explicit medic confirmation. The expectant (black) categorization cannot be AI-assigned at all; it is medic-only, enforced in code.