PHANTOM IS IN ACTIVE DEVELOPMENT. FOLLOW THE BUILD. TRACK DEVELOPMENT
Intelligence Layer

MERCURY AI

From post-flight analysis engine to airborne cognitive system. Real-time 3D reconstruction, edge inference, swarm coordination and predictive environmental modeling — thinking at thirty thousand feet.

200+ TOPS, Target Compute
<1s Target Decision Loop
99% Detection Precision, Benchmark Set
In Design Aerospace Compute Module
Onboard Compute

Built for the Sky, Not Borrowed From the Ground

Mercury runs today on a ground workstation, processing footage after the aircraft lands. The plan is a custom compute module, designed for the heat, vibration, and power limits of actual flight instead of adapted from a desktop PC.

Specification
Dev Rig (Today)
Flight Module (Target)
Platform
Ground workstation
Custom aerospace module
Form Factor
Standard PC chassis
Sealed, vibration-isolated
Power Draw
150W
Under 20W, target
AI Performance
About 100 TOPS
200+ TOPS, target
Deployment
Post-flight only
Real-time, airborne, target
Processing Mode
Batch, video file
Streaming inference, target
Latency
20 to 40 minutes
Under 1 second, target
// compute migration plan
$ plan --target flight_module --build mercury_v2
→ model compilation: FP32 to INT8 quantization
→ expected memory reduction: 4x
→ expected accuracy loss: under 1%
→ architecture shift: batch to frame-by-frame inference
$ roadmap --phase compute
→ months 1 to 2: ground bring-up on the new module
→ months 3 to 4: first airborne integration attempt
→ months 5 to 6: flight testing and latency measurement
_

None of the above is flight-tested yet. It's the plan we're building toward.

Reconstruction

Seeing While Flying, Not After Landing

Standard structure-from-motion waits until you're on the ground to build a 3D model. Mercury's design goal is to build it frame by frame, while the aircraft is still in the air.

100Hz
Local Layer, Visual-Inertial Odometry
Gyroscope, accelerometer, and optical flow fused for fast position tracking. No 3D model here, just where the aircraft is, robust to GPS dropout.
High Rate
2Hz
Window Layer, Sparse 3D Structure
The last 60 frames are kept in active memory, building a rough 3D structure of whatever's immediately around the aircraft, for obstacle awareness.
Active Window
0.1Hz
Global Layer, Persistent Map
A background process folds window results into one persistent map, recognizes places it's already seen, and corrects drift over time.
Background
2Hz
Render Layer, Live Operator View
The current 3D view streams down to the ground station, so an operator sees roughly what the aircraft sees, not a video feed running minutes behind.
Streaming

This architecture runs today in post-flight processing. Live, in-flight streaming is the target for the edge migration phase below.

Object Detection

From Pixels to Something Worth Flagging

A YOLOv26 backbone, compiled for our target hardware and fine-tuned on environmental targets. On our benchmark set, single-frame detection sits at 91% precision, and combining detections across several frames pushes that to 99%.

Camera Frame
640x640px input, 30 FPS onboard
YOLOv26
Compiled INT8, 91% precision baseline
3D Projection
Camera ray through the 3D mesh to a GPS coordinate
Temporal Fusion
2-second tracking window, 99% precision
Alert & Action
Flags an operator or triggers a response rule
TARGETS
What It Looks For
Waste piles, smoke plumes, unauthorized persons, off-road vehicles, and water discoloration are the current training targets.
PRECISION
Where the Numbers Come From
91% single-frame and 99% fused precision are both measured on our own recorded test footage, not flight-validated results yet.
RESPONSE
What Happens After a Detection
Planned behavior: a waste pile drops altitude to document it, a person in a restricted zone gets tracked, a smoke plume triggers a vertical profile scan. An operator stays in the loop.
Swarm Coordination

One Mind, Many Bodies

The long-term design: multiple aircraft sharing one world model, splitting up an area between them, and re-organizing automatically if one drops out. This is a modeled projection based on coverage math, not something we've flown yet.

PHANTOM-01 P-02 P-03 P-04 P-05
Communication Architecture, Target
Frequency band900MHz / 2.4GHz
Throughput500 kbps
Air-to-air range5 km
TopologySelf-healing mesh
Modeled 5-unit coverage12 min / 10km, vs 45 min solo
Predictive Modeling

From Observation to a Recommendation

The goal isn't just a detection, it's a next step. Mercury is designed to model fire spread, pollution transport, and infrastructure wear, and hand an operator a specific, time-stamped suggestion. The operator still decides.

WILDFIRE
Spread Modeling
Fire perimeter, fuel type, terrain slope, and wind feed a coupled atmosphere-fire model, aiming for a spread probability map and firebreak suggestions.
EMISSIONS
Plume Transport
A Gaussian plume model using measured wind profiles, estimating source location and downwind impact to support evacuation planning.
INFRASTRUCTURE
Degradation Forecast
Repeat surveys feed a trend model tracking erosion or pipeline exposure over time, aiming to flag issues before they become failures.
EXAMPLE OUTPUT, HUMAN AUTHORITY PRESERVED

"Deploy containment team to coordinates X, Y. Evacuate zone Z within 45 minutes. Inspect pipeline segment P within 72 hours." The operator approves, edits, or overrides every one of these.

Roadmap

Timeline to Airborne Cognition

01
Foundation
Current
Post-flight 3D reconstruction and YOLOv26 detection running on recorded footage on a ground workstation.
In Development
02
Edge Migration
Months 1 to 6
Move to the custom flight compute module. Optimize the inference pipeline. Ground and airborne integration testing.
Next Up
03
Real-Time Intel
Year 1 to 2
Streaming 3D reconstruction and live detection in flight, with an operator interface that updates as the aircraft flies.
Planned
04
Fleet Cognition
Year 2 to 3
Multiple aircraft sharing what they learn. Swarm coordination and predictive modeling running together.
Long-Term
Join Us

Build the Future with Us

We're always looking for passionate builders, thinkers, and operators to help push autonomous flight forward.

Apply Now
Get in Touch

LET'S CONNECT

Have a question, an idea, or just want to know more about Phantom? Whether it's a technical doubt, a collaboration idea, or you just want to say hi, we'd love to hear from you. Shoot us a message, we're always happy to help.

LinkedIn X YouTube Instagram Email Us