CelerFama
taylortjohnson.com
Autonomous cyber-physical systems

Verified
Physical AI

We design autonomous systems — aerial and ground robots, and the embedded stack that runs them — using agentic AI across the whole design cycle, with formal verification as the evidence that the result is safe.

A quadcopter enclosed by a computed reachability envelope with sensor perception cones
The premise

Building a new robot normally takes a large team — mechanical, electrical, RF, software, controls, safety, supply chain. We think a small team using agentic AI can do it, provided the design is verified rather than merely tested.

How it works

Specify, generate, verify — then go around again

Speed only counts if the result holds up. We treat verification as the inner loop of design rather than a gate at the end, so each iteration produces evidence alongside the artifact.

01

Specify

Requirements and safety properties are written down formally at the start — the operating envelope, the failure conditions, what the system must never do.

02

Generate

Agentic AI works the full stack: board layout, mechanical CAD, firmware, control laws, and the integration between them — the disciplines that usually need separate teams.

03

Verify

Formal verification, reachability analysis, and model-based simulation check the candidate against the specification. What fails comes straight back into the next iteration.

The loop is the point — iterations are cheap when the check is automated
Agentic engineering

One loop across every discipline

Board layout, CAD, firmware, and control design are usually separate tools owned by separate specialists, and much of the schedule goes into moving work between them. We drive them as a single automated pipeline, so a design change propagates through the whole stack in one pass.

  • PCB layout and electrical design
  • Mechanical CAD and structural models
  • Embedded firmware and flight software
  • Control synthesis and tuning
A circuit board, a CAD assembly and software components linked into one connected design pipeline
Assurance

Evidence that the autonomy is safe

Modern autonomy runs on learned components, and a learned component cannot be assured by testing alone. Our foundation is neural network verification and the analysis of neural network control systems — establishing what a system can and cannot do across a whole range of conditions, not only the ones that were tried.

  • Neural network verification
  • Neural network control system analysis
  • Reachability and set-based methods
  • Model-based design and simulation
A ground robot and an aerial vehicle inside a bounded safe-operating envelope with an abstract neural network graph
Where it applies

Platforms built where they are used

We are designing cyber-physical platforms — quadcopters among them — to be produced in North America, with the supply chain treated as a design constraint from the first iteration rather than a problem discovered at the end.

01

Education

Platforms that are affordable to buy, safe to operate, and open enough to teach with.

02

Research

Instrumented, reproducible testbeds for autonomy and controls work.

03

Defense

Trusted, domestically produced systems with assurance evidence built in.

04

Commercial

Inspection, survey, and logistics platforms adapted to a specific job.