Local-first drone footage intelligence

Your footage, searchable.
Your data, never leaving the building.

DroneOS turns hours of aerial video into searchable intelligence — on your own hardware, with no cloud dependency and no footage ever leaving your control.

  Runs fully local · Windows / macOS / Linux · Air-gap ready · Zero telemetry
One click to the moment

Jump to the second.
Watch the AI watch with you.

Live product capture DroneOS player: live 'boat 58%' detection box over playing video, GPS coordinates burned in, location flight track, mission details panel
02 — Player

A live detection box — boat, 58% confidence — rendered over playing 60 fps footage, with GPS coordinates burned into the frame, the flight track on the location panel, and operator notes alongside. Toggle boxes, adjust the threshold, export a clip ±5 seconds, or share a deep link that lands a colleague at this exact second.

Operator truth, on every asset

AI describes.
Operators annotate. Humans win.

Live product capture DroneOS Edit Video Details: display name, location 'North Perimeter, Site A', operator notes, tags, and GPS track pinning
Live product capture Timestamped AI scene log with detection chips: '15s — boat — A solitary boat floats on the water under clear daytime skies'
03 — Metadata & scene log

Every asset carries operator metadata: a display name, a location label like "North Perimeter, Site A", mission notes, tags, and a GPS track — pin one point for the whole video or scrub and drop waypoints so the marker moves. Beside it, the AI writes a timestamped scene log for every segment; analysts can correct any line, and corrections are re-embedded into the search index so human truth beats AI guess.

How it works

Ingest. Search. Act.

01

Ingest

Drop in a mission or point at a memory card. DroneOS preserves an immutable original with a SHA-256 hash computed during copy — chain of custody starts at ingest — then transcodes, extracts keyframes, runs GPU detection, and indexes everything for semantic search. Jobs are persistent and resumable.

FILES SHA-256HASH DETECT+EMBED INDEX
65 files · ~32 GB 4K DJI · ingested faster than realtime*
02

Search

Plain-language queries hit a hybrid engine: semantic similarity plus keyword and object matching, fused, then re-ranked by a local 14B LLM that weighs actions and relationships — not just objects. Search never hard-fails; if a model is unavailable, results degrade gracefully instead of going dark.

SEMANTIC KEYWORD/OBJ FUSE LLM RERANK → RANKED MOMENTS, IN SECONDS
Warm search round-trip: seconds, LLM re-rank included*
03

Act

Jump to the moment. Share a deep link that lands a colleague at the same second. Export a clip. Generate a mission PDF — stats, detection summary, asset inventory with SHA-256 hashes, keyframe highlights with analyst notes. Save any query as a watchlist: new footage is auto-scored on ingest, matches raise alerts for human review.

t=15 CLIP±5 SEC PDF0.5 MB ALERT✓ / ✗
65-video mission report generated in seconds*
* Reference hardware: AMD Ryzen AI Max+ 395 laptop · Radeon 8060S · 64 GB RAM
Mission geography

Every asset labeled.
Every mission mapped.

Live product capture DroneOS mission 'Site A — Perimeter Patrol': 65 videos with labeled assets — North Perimeter, East Gate, Substation 12 — plus search-this-mission and export report
04 — Missions

Each ingestion becomes a mission — here, Site A — Perimeter Patrol: 65 videos, 2h 18m — browsable, searchable within the mission, and exportable as a report in one click. Every asset carries its operator label: North Perimeter, East Gate, Substation 12 — Transmission Corridor.

Live product capture DroneOS mission map: geolocated site pins across the United States for the Site A — Perimeter Patrol mission
Live product capture Labeled infrastructure and public-safety assets: Pipeline Segment 7 — River Crossing, Flood Response — Riverside District, Sector 9 — Access Road Patrol, Search Grid C — Ridgeline Sector, Transmission Line 4B — Vegetation Survey
05 — Map & labeled archive

DJI flight-log sidecars are parsed automatically, and every video is pinned on the mission map — the live archive spans eight geolocated sites across three markets: defense (Sector 9 — Access Road Patrol), critical infrastructure (Pipeline Segment 7 — River Crossing, Transmission Line 4B — Vegetation Survey), and public safety (Search Grid C — Ridgeline Sector, Flood Response — Riverside District).

Measured, not promised

Real numbers from real footage.

1.02×
Faster than realtime — full ingest of a 65-file, ~32 GB 4K DJI mission, detection + indexing included
18×
Faster keyframe extraction vs. naive decoding
14×
Faster image embeddings with fp16 GPU (33 → 2.4 ms per image)
SECONDS
Warm search round-trip, local LLM re-ranking included
0.5MB
Full 65-video mission PDF, generated in seconds
All figures measured on reference hardware: AMD Ryzen AI Max+ 395 laptop · Radeon 8060S · 64 GB RAM
Standing queries

Watchlists work
while you sleep.

Live product capture DroneOS Alerts: three saved watchlists — person near the perimeter, vehicles entering the area, person carrying a backpack — with the banner 'all AI warnings require human review'
06 — Watchlists & alerts

Save any search as a standing watchlist — "person near the perimeter," "vehicles entering the area," "person carrying a backpack". Every new video is auto-scored against them on ingest, and matches raise alerts for an analyst to approve or dismiss. The rule is printed at the top of the screen: all AI warnings require human review.

Evidence-grade output

One click.
One mission report.

Stats, detection summary, a hash-verified asset inventory, and keyframe highlights with analyst notes — a full 65-video mission compiled into a compact PDF in seconds. Every asset carries the SHA-256 computed at ingest, so the report is a chain-of-custody document, not just a summary.

Privacy & deployment

Local-first isn't a feature.
It's the architecture.

DroneOS runs as a single process on your hardware. The backend binds to localhost only. Zero telemetry, no CDN dependencies, and a fully pinned software supply chain with automated drift detection in CI. Local Only mode keeps every AI operation on-device, with your GPU auto-detected — the captures above are running vision and a 14B LLM entirely on an AMD laptop GPU.

Designed for air-gapped, local-first operation; formal certifications are on the roadmap.

Windows · NVIDIA CUDA Windows · AMD ROCm + Vulkan Windows · DirectML macOS · Apple Metal Linux · CUDA / ROCm / Docker Any PC · CPU fallback

Windows installer: per-user, no admin rights required.

Public roadmap · Labs previews · not shipping today

The foundation is live.
Here's what builds on top of it.

Ingestion, AI detection, search, video, and maps are shipping now — everything above is a capture of the running product. These modules are in-app Labs previews, labeled "coming in a later phase."

Roadmap

Intelligence

Cross-mission patterns, predictive alerts, activity heatmaps.

Roadmap

Digital Twin

Orthomosaics, 3D models, before/after, GIS & ATAK export.

Roadmap

Reports Hub

Batch export, daily briefs, evidence packages.

Roadmap

Autonomous Tasking

Coverage-gap suggestions for the next flight.

Roadmap

Tracking & Live

Object re-ID, live RTMP/RTSP ingest, privacy redaction.

See it run on real footage

The demo isn't
a slide deck.

Everything on this page is the real product, captured live: a 65-video mission — over two hours of genuine 4K drone footage across eight sites — searched, mapped, and reported on a laptop. With the network off.

Request a demo
NETWORK: DISCONNECTED  ·  DRONEOS: RUNNING