Nothing on the page matches that.
Which is a good reason to just say it in your own words.
Tell us what you want to buildCodestead · Ogden, Utah
Codestead turns business insight into software built around how your business runs. One senior engineer defines it, builds it, and hands it over. The terms are published, so you read them before we start.
Or press CtrlK to search anything on this page.
01The offer
Custom software for companies that don't have an engineering team and don't intend to hire one. Internal systems, integrations, data and decision tools, focused AI capabilities, and customer-facing products.
You don't need a spec to start. You know the workflow and the customers, and you know where the money quietly leaks out. That's the part nobody can hand you. We bring the product judgment and the engineering, and turn what you already know into something buildable.
The label on the work matters less than where it points, and in practice it points one of two directions. Neither of them starts with the technology.
Direction one
Software that makes the operation faster, cheaper, or harder to copy. Workflow tools shaped to how the work happens on the ground. Integrations between systems that were never designed to talk to each other. Reporting that answers the question you asked. It stays in-house, and it compounds.
Direction two
Software your customers use. Portals, vertical applications, SaaS products, new platforms — usually built on expertise you already have and currently deliver by hand. It leaves the building, and what you currently deliver by hand becomes something you sell.
If a process change, a better use of a tool you already pay for, or something off the shelf solves it, we say so in the first conversation. That answer costs us a project. It's still the right answer, and you'll remember who gave it to you.
02What we've shipped
This is usually where a case study goes. Here are two products instead, built and launched and run end to end.
Both of these are my own companies. They're here to show you the ceiling. Most client work is deliberately simpler than either of them.
Live
Blueprints in, priced estimate out.
A contractor uploads a set of construction blueprint PDFs. The system works out which pages are floor plans and which are elevations or schedules, finds the drawing scale, traces the walls and rooms, reads the finish, door and window schedules, and produces measured quantities — wall square footage, ceiling area, linear feet of base and crown, door and window counts. Then it prices those quantities against the contractor's own rate document. A reviewer corrects anything wrong in a blueprint viewer and exports CSV, XLSX or PDF. Three estimate types are built: paint, drywall, and window and door.
A floor-plan segmentation service and a layout-aware OCR service, both running on hardware we own, both metered as first-class cost providers. The roughly five thousand lines that turn a raw segmentation mask into scale-calibrated, priceable geometry are ours.
Get the scale wrong and every number after it is wrong, so four independent methods run against every sheet. The printed scale bar is measured with Hough transforms and tick-peak detection. The scale notes in the title block are read and parsed. Line primitives and dimension text come straight out of the PDF's path objects, and pixels-per-foot is fitted with RANSAC. Wall-thickness probes from the segmentation model are clustered and variance-checked. Roughly 4,800 lines across four modules, doing the same job four different ways and arguing about the answer.
Each wall segment is scanned bidirectionally to the true wall edge and corners are rebuilt at edge intersections, so concave and L-shaped rooms come out right and doorways stay open. Rooms nested inside rooms become geometric holes, so ceiling area is never double-counted.
Each page is embedded as image and text layer in a single 3,072-dimension vector, and layout-aware OCR chunks share that vector space, so one query reaches both the drawing and the spec text. Roughly ninety hand-authored, domain-expert queries across about thirty task types — including a dedicated set for whether partitions run to the deck or stop at the acoustical ceiling grid.
Every model citation is checked against the retrieved evidence. An unresolvable citation triggers a structured retry; a second failure fails the branch. Citations are then re-grounded to source pixels with a deterministic IoU pass that maps the visual region back to the OCR chunk underneath it — no second model call.
Each customer's pricing logic is compiled into a sandboxed parametric Python function generated from their own rate document and executed under RestrictedPython with a ten-second cap — including a hand-written mitigation for a real CPython str.format sandbox-escape vector.
An automated sweep walks every project-scoped route at runtime and asserts that cross-organisation access is refused. New endpoints are covered without anyone remembering to update a test.
~490k
lines of first-party code
3,064
commits, about twelve months
4,331
automated Python tests
1,322
frontend tests
317
migrations across 91 models
26
services, on owned hardware
These are counts of what's in the repository. We're not claiming a pass rate.
Launching soon
Not deployed yet. No users on it, and nothing to log into.
CAD you drive by conversation. Your model stays a readable script you can open in any text editor.
You describe a part in plain language, or click a face or edge in the 3D viewer, and a tool-calling agent writes or edits the script that builds it. The engine executes that script in a sandbox and renders the result. You iterate by conversation. There's no modeling toolbar. It's built on the OpenCascade geometry kernel, so the output is real B-rep solids with exact surfaces, and the file you walk away with is a script a person can read.
Drag the part to turn it, or move a number and watch it rebuild. Point at a face and the line that made it lights up. Delineo runs this on OpenCascade; a kernel that size has no business on a landing page, so the part here is evaluated by a small one written for the surfaces it needs — planes, cylinders, a cone, one torus.
Point at a face and the line that made it lights up.
flange.pythe model
7 of these lines make faces. Hover one, or tab in and use the arrow keys.
The still above was rendered from this script when the site was built.
Pick any face or edge in the viewer and the engine fingerprints that solid entity, synthesizes a reference expression for it, and round-trip validates it by topological identity before the model ever sees it. The AI is handed a reference that has already been proved to resolve.
When a later edit changes the geometry a reference pointed at, it raises an error and stops. A quiet edit on the wrong feature would surface three revisions later, when it costs a great deal more to unpick.
STEP files are re-imported and compared on volume, bounding box and solid count. STL and 3MF are verified watertight: every edge shared by exactly two triangles, consistently wound. A silently wrong export is worse than a failed one, because you find out at the shop.
No network access, a read-only filesystem, an eight-module import allowlist, and hard CPU, memory and wall-clock limits. Nothing reaches your model unless it executes, produces a valid solid, and has positive volume.
Twenty-six instrumented modeling operations record provenance as the script runs, and every revision shows the material it added and removed, as solids.
Live section planes are cut on the GPU with properly capped solids via a stencil pass — no cap geometry, and the solid count does not matter.
1,180
automated Python tests
359
frontend unit tests
52
end-to-end journeys
Repository counts again, same caveat. Assemblies are first-class too: named instances with explicit transforms, live re-render when a part changes, a bill of materials with CSV export, exploded view, and pairwise interference checking.
And for someone else's business
A gravel company was paying staff to hand-key every weigh-ticket off their scale software into QuickBooks. Copy, paste, ticket by ticket, load after load, every day. We connected the two systems directly. Ticket data lands in QuickBooks on its own now, and nobody re-types anything.
The integration itself was routine. What took the work was understanding the yard from the operator's side and shaping the software around the real process, messy parts included. The same product thinking applies whether the result stays inside the business or becomes something customers buy.
Before
After
hrs →0
daily ticket entry
0
tickets re-keyed
1:1
fit to the real workflow
Other client engagements are covered by non-disclosure agreements and stay off this site entirely, not even anonymised. Industry plus workflow plus timing can identify a client to anyone who knows the market.
03The terms
So there is one set of terms, it is the same for every client, and it is published in full for you to read before you ever get in touch. Codestead keeps the copyright. You get the code and a licence to do essentially anything with it, and once the build is paid for nobody can take that away.
What you get
Here's the whole list, in front of you before you commit to anything. Nobody should be discovering their own terms in year two.
None of this is a concession we make when someone pushes. It's the arrangement we'd want if we were the ones paying, and it's why a business can hand core operations to a firm this small.
What we keep
Codestead keeps the copyright and stays free to build similar software for anyone. Most firms leave you to find that out later, so it is written in bold in the licence and it is written here. What it buys you is a build priced as though it is one of several, rather than one that has to pay for itself on its own.
What you give up
Exclusivity. The code stays ours, and the same capability can be built for someone else in your market unless you pay us not to.
What stays yours alone
Your data and anything you tell us in confidence, always, including in disguised form. Plus any component named in your agreement, which becomes your property outright.
That leaves three things to settle in the same step-two conversation, and only three: the price, which components are named as exclusively yours, and whether you want us restrained from building in your market at all. That last one is available and it is priced, because it is us agreeing to turn down work. Everything else is the same licence everyone else gets.
Who's doing the work
There's no account team here, and no handoff to someone you never met at the pitch. That has a real limit: a firm this size takes on fewer projects at once. We made that trade on purpose, and what you get for it is that nothing about your build ever gets explained to you second-hand.
“I started Codestead because the best software ideas I kept running into were coming from people running businesses — not from software companies. Most of them never got built.”
“I work on every project myself: the first conversation, the scoping, and the engineering. That is deliberate. A firm this size cannot hide a weak build behind an account team, and it means the person telling you what something will take is the person who has to go do it.”
How a project runs
No discovery phase that bills for six weeks and produces a slide deck. Each step ends in something you can hold: a written summary, a priced scope, working software, and finally the keys.
The idea, the bottleneck, or the advantage you can see from inside the business. It doesn't have to arrive as a technical spec. You get an honest read on whether software is the right answer, including when it isn't.
One conversation, plus a short written summary
We pressure-test the value, map the real users and the real workflow, and scope the smallest dependable version worth building. You get a price, a sequence, and an explicit list of what got cut and why.
One to three weeks
Senior engineering in visible increments, with security and reliability sized to the data you're holding and the risk you're carrying. You get working software you can open and use yourself at each stage.
Most first builds run a few months
Deploy into the business, or out to your customers. Then hand it over: the running system, the source, and the documentation.
Support continues as long as you want it, and never becomes a condition of keeping the software
04What this costs
Nobody can price a scope they haven't seen, so we're not going to put a number on this page and pretend it means something.
Step two ends with a scoped first build, and you decide with the price, the sequence and the cut list all in front of you. Deciding no is a perfectly good outcome of that conversation.
Starting is smaller than people expect. Bring us the opportunity and we'll help define the right first build, in one conversation, with no spec required.
The five that come up most, answered in a sentence each.
Say it in plain language. You'll get a straight answer about whether it's worth building, and what it would take.
Or write to [email protected].