Hello. I am the Aegis-X Assistant. How can I help you?
AI VISION PLATFORM · SUPERPIXIS
Aegis-Fire — Predictive On-Premise Vision AI
4-axis fusion vision AI · from fire early-detection to an open platform
A Korean AI fire-detection system that quadruple-verifies the 'motion' of fire from CCTV alone — structurally removing false alarms while catching cigarette-sized sparks and distant fires in real time. Fully on-premise (video never leaves the site) · zero subscription · runs from a single executable on NVIDIA, AMD, Intel, and Korean NPUs.
92.0 / 92.7%
Distant / small fire recall
8.3% → 3.8%
Night-time false alarms (internal)
58.9 fps
1080p real-time (internal measurement)
5 backends
Single model · vendor-independent
Three unsolved market problems
① False-alarm flood
Sunsets, headlights, welding sparks, red signs mistaken for fire → operator alert fatigue → loss of trust.
② Small / distant misses
Cigarette embers, initial ignition, and distant fire points are missed — the golden time is lost.
③ Vendor lock-in
NVIDIA-only → Korean accelerators and low-power edge sites are excluded.
4-axis fusion detection — judged by the 'motion' of fire, not color
Not color or shape, but the 'motion' of fire (rising, flickering, spreading) cross-verified over time — fewer false alarms, no missed fires, minimal idle cost.
Spatial · the AI's eye
Scans for flame and smoke in real time like a human eye. The gatekeeper that selects only 'first candidates' for the quad-check.
Dynamics · motion
Measures the distinctive rising flicker of fire. Filters out 'non-fire light' like signs and headlights, structurally blocking false alarms.
Temporal · time flow
Confirms the 'spreading flow' over several seconds. Only real fire is accepted, not fleeting glints.
Rules · micro heat source
Catches even cigarette-sized sparks by color and temperature. Secures the 'golden time' before a small fire spreads.
Fusion gate · cost structure
Multiplies the 4 axes for a final verdict. Detailed analysis runs only when a spatial candidate exists → minimal idle compute cost, high-margin structure.
Performance — internal measurement
Small / distant fire recall (specialized-training effect)
Distant fire (<1% of frame) 83.3% → 92.0%, small fire (1~3%) 84.0% → 92.7% — +8.7%p each. Cross-validated on unseen real fire footage + public sets.
False-alarm suppression
Night-time urban false alarms 8.3% → 3.8% (−54%) · 0 false positives out of 4 synthetic non-fire samples · 3/3 real fires retained. (Sample sizes are small — the improvement delta is more meaningful than the absolute figures.)
Deep-learning model · training strategy
YOLOv10s ultra-light (~7.2M) · single-ONNX deployment. Focused oversampling of small/distant fires + hard-negative re-training of false-alarm scenes + public/synthetic data diversity + strict train/val separation (no leakage).
Vendor-independent — single model · auto-selected 5 backends
One model auto-selects the optimal accelerator at install (NVIDIA · AMD · Intel · Korean NPU · CPU).
NVIDIA RTX
TensorRT (.engine, FP16)
8.6ms · ≈116fps
Intel Core Ultra
OpenVINO NPU (AI Boost)
~10~20ms · low power
AMD AI Pro
DirectML (DX12)
stable acceleration
No GPU
CPU fallback
always runs
Validation · Internal Measurements
Scores were measured only on unseen real footage (KISA RS / FASDD / real fires); synthetic data was used for training support only. All figures below are internal measurements, not third-party certification results.
11/11
Real-fire clips detected
20/20
Robust to quality perturbation (20 types)
4 → 0
Synthetic false positives (n=4)
3.9%
Normal-footage FPR (t=0.31)
Internal measurement using KISA certification criteria
| Eval set | Accuracy | Recall | F1 | FP rate |
|---|---|---|---|---|
| KISA standard set | 97.8% | 100% | 0.98 | 5.2% |
| Distant fire (<1%) | 94.6% | 92.7% | 0.93 | 4.3% |
| Small fire (1~3%) | 95.6% | 95.3% | 0.94 | 4.3% |
| FASDD distant | 98.0% | 100% | 0.98 | 4.0% |
| FASDD small | 99.0% | 99.3% | 0.99 | 1.3% |
Deployed model merged15 · threshold t=0.30 · measured internally by GaRangBi — not conducted or certified by KISA. Applying KISA's criterion (F1 ≥ 90%) internally yields 0.93–0.99; the false-positive rate on real footage (1.3–5.2%) does not yet meet the certification bar (<1%), so the official test will be re-measured on certification-spec footage. Per-set sample sizes are available on request.
Target segments — B2G demand created by regulation
Logistics centers · warehouses
Early capture of distant fires and smoldering in large spaces (core target).
Municipal integrated control
Multi-channel CCTV fire monitoring · false-alarm reduction eases operator fatigue.
Factories · industrial sites
Electrical fire (arc) and equipment-ignition monitoring.
Public · critical facilities (B2G)
Public and critical facilities — schools, tunnels, parking structures. Public-procurement track engagement is in preparation.
From closed product to open platform (roadmap)
Opening the Aegis Vision AI Core (4-axis fusion engine, 5-way backend, edge on-premise inference) is on our roadmap — Open SDK/API, third-party detection-model marketplace, standard connectors (VMS/IoT), multi-tenant monitoring. These are planned, not currently available; timelines are shared on request.
Aegis brand family
Under the Aegis master brand — Aegis-Fire (fire early detection and monitoring, flagship) and Aegis-X (multimodal data). Aegis-Safety (industrial PPE / restricted zones) is a planned extension. Reusing one vision core across domains lowers marginal cost.
Deployment · PoC inquiry
Aegis-Fire is preparing for KISA certification and has not yet obtained it. For B2G adoption, PoC or partnership enquiries, contact [email protected].
Deployment · PoC inquiryAll quantitative figures on this page are GaRangBi AI Tech internal measurements, not certification or validation results from KISA or any other third party. Official KISA test values will be re-measured on certification-spec footage and may differ. Per-set sample sizes and measurement conditions are available on request.