HP EDGE AI SJSU HACK · SEPTEMBER 2026
Wildfire Edge Sentinel
Context-aware wildfire detection that runs entirely at the edge —
no cloud dependency, no false-alarm flood.
Runs on 1× HP ZGX Nano · NVIDIA GB10 · 128GB unified memory
SSH-only edge node — video and inference never leave the device.
Team: Savitha Vijayarangan · Kruthika Virupakshappa · Sai Chaitanya Munagala · Bavishna Ashok Kumar · Elsa Rose
THE PROBLEM
No connectivity
Remote wildland camera towers sit in weak or no-signal areas, streaming video to the cloud for analysis simply isn't possible.
A hard cost ceiling
Where a link does exist, it's satellite or cellular, billed per MB. Continuous video upload at 24/7 frame rates is not affordable.
A false-alarm flood
Naive smoke detectors can't tell a campfire, a BBQ, or chimney smoke from a wildfire and dispatch drowns in noise.
Illustrative 4-tower deployment × 2 fps — what a naive cloud pipeline would need to move, store, and pay for each day.
Wildfire Edge Sentinel
HP Edge AI SJSUHack
ARCHITECTURE
A cascade on one HP ZGX Nano
Tower & mobile cameras
every frame, on the tower LAN
YOLO11 detector
every frame · ~40 ms
Persistence gate
smoke in 3 frames in a row
Qwen2.5-VL-7B + LoRA
once per event · ~5.5 s · JSON
Severity rules
IGNORE · LOG · MONITOR (re-check) · ALERT
Durable outbox
only ALERTs · waits while the link is down
Team phone + dispatch
ntfy push · incident map · bearings · wind
Wildfire Edge Sentinel
HP Edge AI SJSUHack
IMPLEMENTATION STACK
One ZGX Nano runs the full detection and decision stack
HARDWARE
HP ZGX Nano, NVIDIA GB10, 128 GB unified memory
DETECTION
YOLO11s runs in-process on the GPU at 960 px
CONTEXT MODEL
Qwen2.5-VL-7B + LoRA served through HP zrt and vLLM
APPLICATION
Python 3.10, PyTorch, FastAPI, Pydantic, and SQLite
Wildfire Edge Sentinel · HP Edge AI SJSUHack
ALERTS
ntfy push to the team phone; optional dispatch endpoint
Data
D-Fire · HPWREN FIgLib (20 recorded fires) · pyro-sdis
HP Edge AI SJSUHack
Three problems we had to solve
Catastrophic forgetting
Training on tower footage dropped photo mAP50 from 0.787 to 0.106. Joint training on both restored 0.748, and 0.718 on towers.
One box, big models
A 32B teacher, fine-tuning and live serving share 128 GB of unified memory. We tuned the server's memory share and served adapters over its socket.
An honest comparison
Cloud-only gets the same model, the same frames and real link speeds, so the edge only wins where it really does.
HP Edge AI SJSUHack
Wildfire Edge Sentinel
BENCHMARKS · MEASURED ON THE NANO
Fine-tuning: before → after
Detector mAP50, D-Fire
YOLO-World zero-shot → joint YOLO11s
before 0.002
after 0.748
Detector mAP50, tower smoke
D-Fire-only model → joint model
before 0.198
after 0.718
VLM source-type agreement
7B base → LoRA from the 32B teacher, 500 crops
before 45.0%
after 73.6%
End-to-end recall, 15 fire clips
25 labelled tower clips; precision 0.65
before 0 / 15
after 11 / 15
Wildfire Edge Sentinel
HP Edge AI SJSUHack
IMPACT
Edge vs cloud-only
Junction Fire recording · 6 towers · 266 frames · the same Qwen2.5-VL-7B on both sides
Measured | Edge | Cloud-only |
VLM calls | 11 | 266 |
Tokens | 3K | 447K |
Video uploaded | 0 B | 23 MB |
Time to decision (rural cell) | 5.4 s | 6.3 s |
During a network outage | decides | blind |
THE KEY MOMENT
No network: the edge keeps deciding.
Edge: detector, local VLM, decision — ALERT queued on the device.
Cloud-only: stops at the network boundary. Blind during the outage.
Link restored → the queued ALERT reaches the team's phone.
CONCLUSION
What the system proves
Wildfire Edge Sentinel
HP Edge AI SJSUHack
The decision happens where the fire is.
fewer tokens · no video leaves the tower · keeps working with no network
Thank you — questions?