Our agents take a specification and run the whole job — geometry, mesh, solve, optimize, verify, datasheet — unattended, inside the solver stack your team already uses. The physics grades every run, so the agents keep getting better at your designs without waiting on a review queue.
Every workflow we encode becomes an asset you own — and every unattended run makes it sharper. The learning compounds inside your walls, not ours.
Built by PhDs in electromagnetics and applied mathematics who write automation code every day.
Bring one workflow — we’ll scope it live · NDA before we see anything · Pilot running in weeks, not quarters
The Compute Is No Longer the Limiting Factor. The Manual Work Is.
Not moonshots — bottlenecks
The best AI use case isn’t the problem your team never cracked. It’s the simulation setup, sweep, and report your engineers repeat for the tenth time this month. We find those workflows and remove them.
Expertise trapped in few hands
Complex EM setups depend on a handful of senior engineers — many nearing retirement. We encode their process as reviewable, versioned workflows any agent, and any junior engineer, can execute with the same rigor. The expertise stays when they go.
Your workflows are an asset
Every workflow your team runs by hand evaporates when the run ends — the judgment behind it lives in someone’s head. Encoded, it runs itself — and every result it produces teaches the next run. See where it compounds ↓
From a production pilot with Millibeam, whose configurable antenna-in-package catalog now runs this way. Unattended does not mean unchecked — the solver scores every candidate against the spec before the run picks a winner. Walk through it below ↓
The Problem, in Engineers’ Own Words
From 80+ interviews with RF and simulation engineers across aerospace, telecom, semiconductors, and medical devices.
“I spent 3 hours manually fixing an asymmetric mesh — then gave up for the day.”
RF engineer, medical-device startup
“Waiting 10 hours and then nothing. That’s an entire day lost.”
Antenna engineer, wireless hardware company
“A lot of the RF experts are baby boomers — we are going to lose a lot of these people.”
Engineering leader, aerospace & defense
What We Do
Getting to a workflow that runs itself is an engagement, not a download. We encode the first one alongside your engineers, prove it against acceptance criteria you set, and leave you something that runs unattended — and improves on its own from there.
Agentic Workflow Development
We take a workflow you already run — spec to simulation to report — and turn it into an AI agent that runs it end to end, unattended, with every step logged and reviewable. Then it keeps running, and keeps getting better at it.
- Automated geometry, meshing, and setup
- Optimization loops with verified results
- Auto-generated customer-facing reports
- Integration with CAD, PLM, and your pipeline
Integration & Deployment
AI automation deployed inside your solver environment and your security perimeter — tested, validated, and maintained as vendor releases ship.
- Enterprise solver environment setup
- On-premises or private deployment
- Security and compliance configuration
- We carry the maintenance burden
AI Enablement for Engineering Teams
Your engineers build the automation themselves — on our platform and domain primitives — so you capture the capability without carrying the maintenance burden alone.
- Hands-on automation workshops
- Build on our solver-integration layer
- Playbooks, documentation, and mentorship
Dedicated RF & EM Engineering
Deep electromagnetic expertise on demand — for complex simulations, wireless system modeling, and trusted subcontracting on third-party projects.
- Antenna, MIMO, and multi-user network studies
- Ray-traced propagation with Ansys SBR+
- OFDMA and link-level KPI extraction (SINR, capacity, BLER)
- Complex simulation troubleshooting & custom algorithms
Product ID In. Optimized Design and Datasheet Out.
Millibeam’s Heaviside line spans WiFi 6, WiFi 7, and 5G antenna-in-package modules, and every customer order arrives with its own band, gain, and matching targets — an engineer-day of expert Ansys HFSS work per part. We encoded their process as an agent-run workflow, and it now takes 1.5 hours of unattended agentic work.
The workflow became a skill — data, not code
Their engineers’ process, written once as a reviewable document the agent follows: load the right catalog model, set targets from the customer spec, optimize only within manufacturer-approved bounds, verify the winner, and generate the datasheet.
One unattended run, end to end
Baseline solve, KPI check against spec, Bayesian optimization over approved parameters, re-solve of the winning candidate, and a customer-facing datasheet — patterns, gain, matching, efficiency — with zero hand-editing. No engineer in the loop, and none needed: the KPIs come out of the solver, so the run can tell for itself which candidate won.
Every run leaves a paper trail
Every solve, variable change, and decision is logged. Each run keeps its request, decision log, and working solver project on disk — replayable and reviewable. Optimization their engineers can defend — and the same record is what the agents learn from, since every logged outcome is a graded example of what worked on their parts.
Where it goes: every customer request, automated
The pilot automates one workflow. The end state is a fleet: every incoming customer configuration handled the same way — engineers set the targets and the acceptance criteria while agents fan out across products, specs, and what-if studies, keeping every solver seat busy and getting better at each part they touch.
Every Run Makes the Next One Better
Simulation has what most AI domains lack: an objective referee. A solve converges or it doesn’t; a design meets the spec or it misses. So the agents can grade their own work and improve on it, with no review queue in the way. And the learning stays where it was earned — AI labs pay six and seven figures for recordings of how engineering teams work, to train models for everyone else. We do the inverse: every workflow we encode, and every logged run it produces, stays yours and compounds inside your walls.
A specification goes in. Geometry, mesh, setup, sweep, optimization, verification, and the customer-facing datasheet come out — an engineer-day of expert solver work, done in about 90 minutes with nobody watching.
The solver is the referee. Convergence, S-parameters, gain, efficiency, margin against spec — objective numbers the agent reads for itself. It does not need an engineer to tell it whether the design met the requirement.
What worked feeds back into the workflow — better starting points, tighter bounds, fewer wasted solves. The tenth run on your parts is not the first run over again.
Every run’s paper trail — requests, decisions, solved projects — is training data you own. It feeds custom AI models built on your simulation data, so your operational knowledge compounds inside your walls instead of being sold outside them.
How We Work
Your team keeps operating normally throughout — about an hour of engineer time per week during a pilot. Your engineers set the targets and the acceptance criteria; the agents do the rest and score themselves against them.
Discovery
Walk us through one workflow — a 30-minute screen share of the work your engineers repeat most. We tell you on the spot whether it’s automatable, what a pilot looks like, and what it would save. Free, under NDA.
Pilot
We embed with your team and encode one workflow you already run — weekly working sessions, about an hour of your engineers’ time per week. You judge the results against your own acceptance criteria.
Deploy
The automation moves into your environment — your security perimeter, your solvers, your data — with testing and validation before anything touches production work.
Scale & Support
From one workflow to a fleet: engineers set direction while agents fan out across products, specs, and what-if studies — each run feeding the next. We maintain it as your solvers evolve.
Why Divergent Physics
Engineers, Not Career Consultants
Founded by three PhDs in electromagnetics and applied mathematics, with industry experience at companies including Apple and Ansys. The people on the call are the people in the code. Meet the team →
Trusted in Production
Beyond the pilot above, we are delivering a milestone-gated automation program for the RF systems team of a Fortune-100 consumer-electronics manufacturer — invoiced against acceptance criteria, not hours. Developed with RF engineers across aerospace, telecom, defense, and advanced hardware.
Built on Our Own Platform
Every engagement is delivered on the agent platform we build and maintain — the RF and EM understanding layer that turns a request into a correct simulation, and turns every finished run into a better next one. Prefer self-serve? Run the same agents yourself.
S11 compared across design variations — generated from a natural-language request against a live HFSS project.
Questions Engineering Leaders Ask First
How can it run unattended and still be trusted?
Because the referee is physics, not a person. Convergence, S-parameters, gain, efficiency, margin against your spec — the same numbers your engineers would check are numbers the agent reads directly out of the solver. A candidate that misses the criteria you set does not get reported as a win. And every solve, variable change, and decision is logged, so any run can be replayed and audited after the fact.
Can this run air-gapped?
Yes. Deployments run on-premises inside your security perimeter, including fully local LLM inference — no design data, geometry, or results leave your network.
Who owns what we build?
You own the workflows, code, and deliverables produced for your designs. We retain the underlying platform. Every engagement is scoped that way in writing, up front.
What happens when the next solver version ships?
We track vendor releases and integration drift as part of the engagement, so the automation keeps working — that maintenance burden is ours, not your engineers’.
How do you charge?
Milestone-gated, fixed-scope phases with acceptance criteria you define. A milestone passes, an invoice releases. No open-ended hourly billing — and for ongoing support, a monthly retainer is available.
Bring Us the Workflow Your Engineers Hate Repeating
In one call we’ll tell you whether it’s automatable, what a pilot looks like, and what it would save you. No slideware — we show you a real workflow running.
Free initial consultation
NDA-friendly from the first call
Pilot scoped in weeks, not quarters
Let's talk about your workflow
Book a 30-minute consultation with an AI engineer — or send your project details below and we'll reply within 24 hours.