AI automation that runs
on your servers, not oursCustom workflows and agents for e-commerce and brand operators
Systems live
7
Running on clients' own servers
2
Longest continuous run
7 months
Data retained in third-party SaaS
0
Who you're dealing with
We didn't learn this on client projects
We are our own first client
Before this was a service, it was our own problem. We run several direct-to-consumer storefronts in 10+ languages, with real orders and real complaints arriving every day. Every system on this site was built to fix something in that business first.
The person you talk to is the person who builds it
No sales layer, no handover halfway through, no “let me check with our engineers.” What you describe in the first call reaches the person implementing it without passing through a translation layer.
We don't subcontract
The most common disappointment in this industry: senior people show up to win the work, juniors show up to do it. That failure mode isn't available to us — the people who deliver are the people who pitched.
The honest version of our track record: we are not the vendor with fifty logos in your industry. We're the one whose own business breaks if this stops working — and three of these systems now run on other companies' servers. More about how we work →
What we build
Five things, and they connect to each other
Storefront operations & support agent
The storefront itself, plus the content, support and multilingual work that keeps it running. Built as one line rather than four disconnected tools.
- Support agent trained on your own tickets and policies — most enquiries resolved without a human
- Blog and product content pipelines where topics come from search gaps and real support tickets
- 10+ languages from one knowledge base, with a shared glossary
- Ad and social data pulled back to actual on-site revenue, not platform self-reporting
Visual & video asset pipeline
Product photography in, a consistent batch of scene images, sized variants and short-form video out — with the product itself kept intact rather than reimagined.
- Three tiers, chosen by how exact the product has to look
- Optional brand-specific model trained on your own product photos, owned by you
- Human selection stays in the loop — nothing publishes unreviewed
Internal knowledge hub & Q&A agent
One place where policies, processes, product facts and past decisions live — and one agent that answers from it, with sources.
- Permissions follow your org chart: restricted content is not retrievable, not hidden after retrieval
- Answers cite their source; when there's no answer it says so and routes to a person
- The same knowledge base feeds internal Q&A, customer support and content
Multi-department transformation
For companies where several departments are stuck on repetitive work. One foundation, seven departments on it — and a platform you use to add the next role yourself.
- Finance, people, engineering, sales, e-commerce, design and store management on one knowledge base and one permission model
- A prompt library organised by role, plus a training session per department, run at your offices
- No-code role blocks: a new role gets its helper without joining an engineering queue
- Delivered department by department — the first one doesn't wait for the rest
A single workflow module
The lowest-commitment way in. One trigger, one process you can describe, one output, one human checkpoint.
- Runs standalone, but with interfaces built in from day one so it can be extended later
- Priced and scoped as one project — you can stop after it
- If it can't be described as one trigger and one output, it isn't a module yet
What changes
The same week, before and after
Handled by the system
- ✓Support enquiries answered from your own policies, around the clock
- ✓Content drafted, translated and queued — a person only signs off
- ✓Product images produced as a consistent batch, not one at a time
- ✓Anomalies surface on their own, with the likely cause attached
Handled by hand
- ✕The same question answered by hand thirty times a day
- ✕One article, one language, one person, one afternoon
- ✕Images outsourced, then waited on for days
- ✕Problems discovered when a customer reports them
Why self-hosted
“Own” is not a marketing word here
It runs where you decide
Your own servers or your own private cloud. Not our infrastructure, not a shared tenant. We don't operate it for you afterwards and we don't keep a back-door account.
You get the source code
It lands in your repository, along with deployment docs and a runbook. We write the platform layer ourselves and hand you the source — the foundation is standard tooling, so there is no black box only we can maintain.
It isn't priced per seat
One build, plus an optional annual maintenance subscription. Growing the team doesn't grow the bill, which is what makes it worth automating high-volume work at all.
Your data doesn't leave your boundary
Where a third-party model is used, we list every call and everything it would send during the proposal stage, and you can veto any of them line by line.
What we don't do
Four things we won't sell you
We don't promise search rankings
What we deliver is a content pipeline and the capacity to run it, not keyword positions. Anyone promising positions is selling something else.
We don't run your ad accounts
We do the homework — data consolidation, attribution against real orders, diagnosis, ad copy and creative, recommendations. The decision to spend money stays with the person accountable for it.
We don't promise a smaller team
This work removes repetitive actions, not roles. Exceptions, judgement calls and accountability don't go anywhere — and those are what actually consume people.
We don't sell a SaaS subscription
There is no hosted product to sign up for. If a ready-made tool genuinely fits your case better, we'll say so during the diagnostic.
Who this works for
We take on a small number of projects, so we're selective
The process is settled
If the rules might be overturned next month, automation just cements a temporary workaround into policy.
The volume is real
Something that happens three times a month won't save enough time to justify the meetings needed to start it.
Someone can decide
If three departments still disagree about the rules, the bottleneck isn't technical and we can't fix it.
The flip side is capacity: if the schedule doesn't have room, we'll tell you that instead of signing first and finding out later.
Where it starts
One diagnostic, credited back in full
$79960–90 minutes, screen share
We walk your existing systems and process together. Suited to a single team or a single workflow that can be demonstrated in the back office.
from $4,9991–2 days on location
We shadow the work where it happens — office, warehouse or production floor. Many bottlenecks are only visible on site and never come up in a meeting. Travel billed at cost.
Single workflow module
from $2,799
You walk away with
A deployment list
Diagnostic fee
Credited in full
Quoted in
USD
Project pricing beyond that depends on scope and how many systems have to be connected, so it's quoted after the diagnostic rather than guessed at beforehand.
Questions
What people ask before the first call
01Where does our data actually live?+
On infrastructure you control — your servers or your private cloud. We don't host it and we don't keep a standing account; the keys stay with you. If a third-party model is involved in a step, that step is listed explicitly in the proposal along with what it would send, and you can reject any of it.
02What happens if you disappear?+
Three things make that survivable, and they're all in place before the last invoice: the source code sits in your repository, the source code — platform layer included — goes to you, the foundation is standard tooling rather than anything only we can maintain, and each phase is delivered as something that works on its own. Another engineer can pick it up without talking to us.
03How long does it take?+
We don't publish a timeframe, because it wouldn't hold. This is custom work: industry, process and the state of your data change the effort by several times over, and how fast it goes depends as much on your side as on ours — whether the data can be reached, and whether someone resolves disagreements when they come up. An average would mislead both of us. What you do get at the end of the diagnostic is a deployment list ordered for your situation: what to do first, which tier it lands in, and how many systems have to be connected. Delivery then runs in blocks, and each finished block is usable before the last one is done.
04Do you work with clients outside China?+
Yes. The work is remote by default and the on-site diagnostic can be arranged where travel makes sense. The systems themselves are language-agnostic — our own storefronts run in 10+ languages, which is where most of that work was proven.
05We already have tools for some of this. Does that rule us out?+
No, and it's the normal case. The usual problem isn't a lack of tools, it's that each tool holds its own copy of the truth and none of them talk. Existing systems get connected rather than replaced — the diagnostic starts by mapping what you already have.
06Will adding something later cost extra?+
Yes, and it's written into the contract at signing rather than raised mid-project. Day-to-day maintenance — changing parameters, adjusting wording, adding knowledge entries, keeping up with model and API versions — your own technical people can handle, or you can buy an annual maintenance subscription. A new agent, a new workflow, or connecting a new business system is new development and is scoped and priced as a new project.
Start with one diagnostic
Tell us what your team does by hand most often and roughly how often it happens. That's enough to say whether this is worth a call — and we'll say so if it isn't.
[email protected]The findings are yours either way — including if you decide to implement them without us.