Agents that finish the job
Software that plans a task, uses tools to carry it out and stops when it should — with step limits, spend caps, approval gates and a full trace of what it did.
AI and software development, built to ship.
We build AI features and the products around them — assistants that answer from your own documents, agents that complete real work, and the web applications they live in. Remote, worldwide, and delivered as working software rather than slide decks.
Most AI projects fail on architecture, not on the model. On whether a workflow or an agent fits the problem, where state lives, what happens when a tool errors, and who approves an action that cannot be undone.
Drawn in code, not photographed: every line here is SVG generated by the same build that produced this page. The screen types, the sketch draws itself, the city flickers.
Three things
we get hired for
Most teams arrive with a problem and a budget, not a spec. The first week is spent turning one into the other, and occasionally concluding you do not need a custom build at all.
Architecture before code
Workflow or agent, where retrieval sits, which model handles which path, what happens when a tool fails, and what it costs at ten times the volume. Decided first, written down, and testable.
The product around it
The web application the AI lives inside: authentication, billing, admin tooling and an interface your team can operate without a manual.
1
Developer on your project
Start to finish. No account manager, no handoff to a team you have not met.
6–10
Weeks, scope to handover
Working software in week one, then iteration against an agreed measure.
14
Day invoice terms
Fixed price agreed in writing first. Third-party costs billed at cost.
2
Weeks on call after launch
Included as standard, alongside a walkthrough for the engineers who inherit it.
Watch one
get built
A chat surface, an agent trace and an evaluation table, each assembled line by line in the order we would actually write them.
Chat.tsx
// streaming chat, top down
function Chat({ messages }) {
return (
<Thread>
<Bubble from="user" />
<Stream tokens={reply} />
<Composer onSend={send} />
</Thread>
)
}
function Chat({ messages }) {
return (
<Thread>
<Bubble from="user" />
<Stream tokens={reply} />
<Composer onSend={send} />
</Thread>
)
}
Preview
Ask me anything about how this gets built.
How fast can you ship a prototype?
A working screen in your hands inside a week.
Send a message↵
Project shapes
01 / 04Support assistantDeflection
Grounded answers from your own help content
Judged on how many tickets resolve without a handoff, measured against your current rate.
Document processingExtraction
Invoices, contracts and forms into structured data
Judged on handling time and how often a person has to correct the output.
Evaluation rescueMeasurement
Working out why an existing system underperforms
Usually retrieval or chunking rather than the model. Judged on a before-and-after suite.
Internal assistantKnowledge
One place to ask across scattered documentation
Permissions respected, every answer showing the source it came from.
What we build with
Boring choices where boring is right, and newer tools only where they earn their place. Everything runs in accounts you own.
InterfaceReactTypeScriptNext.jsCSS a human can readdaily
MotionFramer MotionCanvasWebGL when it earns itdaily
StreamingServer-sent eventsPartial renderingResumable runsdaily
Model sideAnthropic SDKOpenAI SDKTool callsRetrievalEvalsweekly
TestingVitestPlaywrightCI on every pushTyped end to endevery project
HostingYour AWS or AzureVercelCloudflareInfrastructure as codeyour account
HandoverStorybookPlaywrightA README your team will readevery project
Nothing here is chosen because it is new. Each has to survive the same three questions: can you hire for it, will it still be maintained in three years, and how expensive is it to leave. Where your team already runs something reasonable, we use that instead.
Scope it in
ten questions
Answer ten and you get a written brief you can copy into an email. It costs you two minutes and saves a week of back-and-forth. Nothing is sent anywhere until you send it.
These are the ten things we need to know before quoting. Answer them in an email if you prefer.
- What do you want to exist that does not exist today?
- What stage are you at?
- Where does the knowledge live?
- Who uses it?
- What does it need to connect to?
- How will you know it works?
- What is the cost of a wrong answer?
- When do you need it live?
- What budget band are you working in?
- Who signs it off?
A project that should not exist is cheapest to stop in week one. If an off-the-shelf tool solves it, we will say so.
Principle 01 — honest scoping
How a
project runs
Six to ten weeks, start to handover. No discovery phase that ends in a slide deck.
01
One week to a working screen
Before any planning, we build the hardest screen with your real model behind it. It's faster to argue with something clickable.
02
We break it on purpose
Slow responses, empty results, a model that gets it wrong. These are the states users actually meet, so they get designed first.
03
Build the real thing
Typed components, tested, in your repo and your design system. No handoff document — we write the code your team keeps.
04
Hand over the keys
A walkthrough with the engineers who inherit it, and two weeks on call after launch.
Questions
we get asked
The five that come up before almost every engagement. The full list, including contracts and confidentiality, is on the FAQ page.
What will this cost?
A scoped feature runs $8,000 to $35,000; a complete product $18,000 to $60,000; a small well-defined build starts at $2,500. The figure is fixed in writing before anything begins, and third-party service costs are billed at cost rather than marked up.
Does our model need to be finished first?
No. We work against whatever exists — a rough endpoint, a notebook, sometimes a person standing in for the model. Waiting for a stable API before designing the interface is how teams discover in month four that the interaction was wrong in month one.
How do we know it is working?
Because the measure is agreed before the build: the metric, the baseline it is measured against, the method, and the bar the system has to clear. If those four cannot be settled, that is a signal the project is not ready to start.
What do you need from us?
Access to a working endpoint, one engineer we can ask questions of, and one person who can decide without a committee. That is genuinely it — research, states and edge cases come from our side.
Who owns what you build?
You do, outright, on full payment — code, prompts, evaluation suites and documentation. Reusable tooling underneath it comes with a perpetual licence, and any open-source components are listed at handover so your team knows the obligations that came with them.
Got a problem
worth automating?
Frequently asked questions
What kind of work do you take on?
Two connected things: AI features (assistants, agents, retrieval systems, automations, document processing) and the software they need around them (web apps, APIs, dashboards, integrations). Most projects are a mix of both, because an AI feature with no product around it rarely reaches users.
Do you work with clients in other countries?
Yes. Work is fully remote and clients are spread across North America, Europe, the Middle East, Australia and Asia. We keep a few hours of overlap with your working day, communicate in writing by default so nothing depends on a meeting, and deliver on your repository and your cloud account.
How quickly can something be live?
A working prototype against your real data usually takes one to two weeks. A production feature with authentication, monitoring, error handling and a proper hand-over is typically four to eight weeks, depending on how many systems it has to touch.
What does it cost?
Small scoped builds start around $2,500. Most production projects land between $8,000 and $35,000. Monthly retainers for ongoing work start at $3,000. Every project gets a fixed number in writing before any work begins — see the pricing page for the detail.
Tell us what you are building.
Send a short description of the problem and we will reply within one business day with an honest view of scope, cost and whether we are the right person for it.
Or email directly: contact@hire-ai-dev.com