An operating business that became a software company by accident
We run a home services business in Seattle — thirteen trucks, crews, a shop, a dispatch board. We got tired of paying for AI that didn't survive contact with a real operation, so we built our own on hardware we own, and the tools turned out to be worth more than the software we'd been buying.
Every product below solved a problem in our own business before it was ever offered to anyone else. That's the whole filter.
Eight records, stage-labelled
The AI Operator
ShippingOne AI employee you talk to, that stands up and runs every other agent in your business.
AI Estate Audit
ShippingFind out what you've actually built before anyone asks you to spend money.
Local AI Deployment & Cost Architecture
ShippingMove the repetitive work off per-token billing and onto hardware you own.
Guardian Layer — Safe AI Outbound
ShippingThe gate that goes in before AI is allowed to talk to a customer.
Fleet Intelligence
ShippingA ranked morning list for the shop, generated overnight for the price of electricity.
Territory Intelligence
ShippingEvery system you own, joined on geography — plus the public data nobody bothers to pull.
Decision Review Panel
ShippingPut the plan in front of independent reviewers who can't see each other's answers.
Field Vision & Operational Robotics Dataset
DevelopingMeasurement today. The training data robotics will need tomorrow.
Stages here match the ones on our Inception product records exactly. Where something is still being built, it says so.
Three theses, one arc
The products above are what runs today. They exist to fund and feed three longer bets, all of which depend on inference we own rather than inference we rent. We are explicit about which of these is revenue and which is a thesis.
Dispatch now, DRIVE later
The route optimisation we run every morning is already the autonomy decision layer. Same optimiser, same territory model, same constraints. The only thing that changes on the day autonomous vehicles are viable in this category is that the route output instructs a vehicle instead of a driver.
Autonomy doesn't arrive as a truck you buy. It arrives as a dispatch brain you already trust. We're building the second thing now so the first one is a swap, not a rebuild.
The training data is already being generated
The race for physical-world training data is on and the labs are paying for it. What most operators do not realise is that they are already generating it — skilled physical work, in real cluttered rooms, photographed every single day. It walks out the door unstructured and unowned.
We grade that photography with vision models today because the measurement pays for itself. The structured record accrues as a by-product. That is what makes the position credible: the capture is funded by something that already works, so the dataset compounds whether or not manipulation policies land on schedule.
Marked Developing on purpose. A robot that clears a garage is not a near-term product, and anyone telling you otherwise is selling something.
The same method, pointed at a different org chart
Everything we built internally came from one exercise: list what each function of a company does, hand the mechanical part to an agent, run it on hardware we own. That exercise is not specific to trucks.
Point it at an independent creator or a one-person innovator and the org chart you take apart is the company that would have signed them — publicity, distribution, licensing, business affairs. Rebuilt as agents on a device in their own studio, on local inference, so the unreleased work never leaves the machine and the ownership never changes hands.
This is the step that turns an operating company's tooling into a platform. It is the earliest of the three and we label it exploratory.
The through-line is the same in all three: inference on hardware the customer owns. Edge decisioning in the field, dataset capture at the source, and generative media rendered locally instead of billed by the token. That is the reason the economics work at our size, and it is the reason the data stays with the people who produced it.
Specific asks, not a general hello
- →Guidance on NIM microservices and the NeMo Agent Toolkit. We're considering both and would rather be told where they fit than guess.
- →A reality check on our vision roadmap. Load grading and scene reconstruction feeding an Omniverse and Isaac training pipeline — we want to know if the sequencing is sane before we spend on it.
- →Introductions in home services and last-mile fleet. This category is under-served and buys slowly, but keeps what it buys for a decade.
- →A pointer to the right autonomy track. Not because we're buying autonomous trucks this year, but because the dispatch layer above should be built to hand over cleanly.
- →The right home for sovereign, customer-owned inference. Our third thesis puts a full agent stack — including generative audio and video — on a device the customer owns rather than in our account. We'd like to know which reference architectures and programmes that belongs to before we commit to a shape.
The long-form version of each of these is a recorded conversation rather than a whitepaper — the actual build discussions, including the parts that didn't work.
See the services