Grounded.
Bounded.
Verified.
Tychon puts AI to work on engineering problems that have to be right. So far that means three space programs, each taken from physics simulation to a model built for the chip. In fields with little public data, an LLM's answer is a guess that reads like a fact, and in physical AI a wrong answer controls the hardware.
METHOD
LLMs do the engineering, with a person directing and checking the work. A small, task-specific model is what ends up in the device.
The LLMs are working on problems their training barely covers, so each use case is grounded in its own reference library, built from the physics, the components, the past work and the code.
A grounded answer can still be wrong. Nothing an LLM produces counts as verified until it has passed tests and simulation, then run on the hardware itself.
Where a model's output drives hardware, the cost of a wrong output is far higher, because the output is a command. So that model is bounded. It only proposes. On the quantum sensor controller a clamp in the chip's logic caps every proposal before it reaches the hardware, and a watchdog catches faults.
WORK
Simulate the world.
Train the model.
Deploy it in the device.
Tychon has run all three as one workflow since 2024, first on a star tracker, then a spectrum monitor and most recently a quantum sensor controller. With each program the LLMs took on more of the engineering, from writing the code on the first to running the FPGA tools and keeping the evidence record on the last.
QUANTUM SENSING
Bounded AI control electronics for NV-diamond magnetometers.
Show the stepsHide the steps
Simulate the worldPhysics simulation of the NV-diamond sensor
Train the modelA 2,905-parameter INT8 CNN that corrects the PID command
Simulate the worldPhysics simulation of the NV-diamond sensor
Train the modelA 2,905-parameter INT8 CNN that corrects the PID command
Deploy it in the deviceProduction FPGA silicon, with a clamp in the FPGA logic that limits the correction to ±4.5% of the actuator range
ResultAbout 53% of the improvement available under the clamp, on held-out data. The fault watchdog was measured at a 2.000 ms median response across 120 induced trips.
VISION
Star tracker for satellite attitude determination.
Show the stepsHide the steps
Simulate the worldA configurable camera simulator, with images checked against real starfields
Train the modelA 3.2-million-parameter MobileUNet, a CNN that finds star centers
Simulate the worldA configurable camera simulator, with images checked against real starfields
Train the modelA 3.2-million-parameter MobileUNet, a CNN that finds star centers
Deploy it in the deviceInference deployed on an RT PolarFire SoC FPGA for a SpaceWERX-funded U.S. Space Force program
ResultRMS centroid error cut from 0.675 pixels with the conventional method to 0.156 pixels, 77% lower.
SPECTRUM MONITORING
Signal classification and interference detection for satellite radio links.
Show the stepsHide the steps
Simulate the worldSimulated radio signals with injected interference
Train the modelA CNN that classifies 11 modulation types from raw I/Q samples, and an autoencoder that flags interference it was never taught
Simulate the worldSimulated radio signals with injected interference
Train the modelA CNN that classifies 11 modulation types from raw I/Q samples, and an autoencoder that flags interference it was never taught
Deploy it in the deviceQuantized to a 45 KB INT8 model for FPGA deployment
ResultClassifier reduced from 455,691 to 28,523 parameters, 94% smaller.
FOUNDER
Jason Phillip builds physical AI for space, most recently the controls for quantum sensors. He was an officer in an Army combat engineer unit and ran a home improvement company for 12 years before switching careers into data science and LLMs in 2022. Since then he's directed AI through the most difficult engineering problems he could find, catching where it fails and turning its output into systems with measurable, externally checkable results. He founded Tychon to do that work.