Trades businesses run on a phone in a truck and a workbook in the office. We build tools that work on both: ordering from the job site, submittals that assemble themselves, quotes the homeowner can see.
They bid large projects in a single Excel workbook: 46 sheets, more than 9,000 formulas, around 1,300 references crossing between sheets. Then the submittal, the document an architect approves before anything can be ordered, was assembled by copying cells out of that workbook and pasting them into Word. Revision tracking was a person typing current values into a column labelled Previous.
What we built
A builder that treats a hardware set as data rather than as a document. Door schedules and hardware sets, glass types, a product data library and a warranty library, all attachable per submittal and clonable from a previous project. The packet generates itself: cover, contents, scope, schedules, hardware sets, and the manufacturer spec sheets appended on the end.
Where it stands
Generated PDFs are deduplicated on a hash of the source data rather than of the file, because the renderer does not produce identical bytes twice from identical input. Legends filter themselves down to the finishes and manufacturers a project actually uses. Pricing deliberately stayed in Excel; the job here was the paperwork nobody went into construction to do.
ConstructionBuilding now
Field materials ordering platform
For an electrical contractor running crews across multiple job sites
The problem
Electricians in the field texted or emailed what they needed to one person at the shop, who then phoned four supply houses one after another to work out who had stock, at what price, and how soon. Nothing was written down as a catalogue, nothing recorded what a job had consumed, and the same part had a different name depending on who was asking for it.
What we built
One request that carries materials and tools together, opened on a phone at seven in the morning and still there on the office desktop at two in the afternoon. A curated catalogue where near-identical parts collapse into a family with a size picker. Trade names anyone can attach, so the thing you call it finds the thing it is. You can also just say the list out loud and have it split into lines.
Where it stands
Per line, it ranks the supply houses by stock, then price, then lead time, and shows where each number came from and when it was last seen. Across a whole order it will split the buy between houses and tell you what that saves. The spoken order is deliberately constrained: the model may only split and label what was said, and every line it produces has to be a literal substring of the transcript, so it cannot invent a part nobody asked for.
ConstructionBuilding now
AI Renovation Visualizer and Quote Tool
For an exterior home renovation contractor
The problem
A rep would meet the homeowner, photograph the house, take notes, then drive back to the office and build the proposal in separate estimating software, emailing a couple of options over the following days. In between, the customer was asked to picture new siding from a handful of small samples and a description.
What we built
Photograph the house on a phone, pick the upgrades, and get a rendered image of the finished result beside the original. The quote is editable on the same page, and it prints as a before-and-after proposal carrying their logo and their address block.
Where it stands
The render happens on the server, so the API key never reaches the browser, and takes somewhere between ten and forty-five seconds while the rep is still standing in the driveway. The proposal was styled to disappear into the paperwork they already hand over, which was the actual brief: if somebody is flipping through his folder, this should look like it belongs there rather than standing out.
