Everyone is selling merchants a chatbot for the shop window. That is not where a growing store loses money. It loses money behind the scenes: stuck orders, returns, chasing suppliers, catalog drift, and a support queue that is the same five questions all day. I build the agents and automations that run that work.
For eCommerce founders and ops leads doing roughly 2,000 to 50,000 orders a month, anywhere in the world. Independent, senior, and the person who builds it is the person you talk to.
Past a few thousand orders a month, almost nothing that hurts is customer-facing. It is repetitive operational judgment, done by people, at volume, all day. Every one of these is a decision with rules behind it, which is exactly what an agent is good at.
Stuck, failed, mismatched, address-broken, out of stock after purchase. Someone opens each one and decides.
Undelivered and returned stock, each needing classification, a decision, and a refund or reship path.
Where is it, why is it late, is the price still valid. The same message, to different people, forever.
Titles, variants, images, stock and pricing drifting out of sync across every channel you sell on.
The same five questions, answerable from data you already hold, absorbing a person's entire day.
Payouts, shipping bills, COD remittance and returns that never quite agree with each other.
You cannot hire your way out of this. Volume goes up, headcount goes up, margin goes down. The fix is a system, not another person.
Built onto the tools you already run. Nothing here asks you to migrate platforms or replace your team. Each one starts as a single workflow and earns the next.
An agent watches the order flow, catches what is stuck or wrong, resolves what has a clear rule, and escalates only the cases that genuinely need a person. Your team stops reading every row and starts reading the exceptions to the exceptions.
Automatic classification of what came back and why, the refund or reship decision applied consistently, and the patterns surfaced. Most RTO is predictable before dispatch, and the same agent can flag it at the point of order.
Product content generated, normalised and kept in sync across channels at volume. Stock, price and variant drift caught automatically instead of by a customer complaint.
The repeated questions answered from your own order and shipping data, not from a generic model guessing. Anything involving money, anger or ambiguity goes to a human, deliberately.
Also: supplier communication, reconciliation, internal reporting that assembles itself, and the pipelines connecting them. If the work is repetitive and rule-shaped, it is in scope.
Client and employer names are withheld deliberately. The work and the numbers are not.
The situation. Sellers running branded dropshipping had to stitch their operation together across a chain of middlemen: sourcing, fulfilment, shipping and branding, with little control and no visibility.
What was built. The platform designed from zero, plus the design system underneath it as a single source of truth, so every surface stayed coherent while the team and feature set scaled hard. The flagship flows were the operational ones: product research, multi-state order management, and merchant onboarding.
The outcome. The system scaled with the business and lifted on-time delivery substantially, which in this category is the whole game.
The situation. Producing bespoke supplier websites was slow and repetitive. The thinking was real work; most of the build was the same job done again.
What was built. A versioned engine that assembles a complete, bespoke site from a single data file. Scrape, generate, live demo, the same day. The design system is the input and the website is the output.
The outcome. Live builds running across four countries. The tool itself was sold to a company, and the same pipeline won a client build contract. This is the clearest example of how I work: find the repetition, then delete it.
The situation. Most agent demos work once and break in production. The difference is not the model. It is whether anyone designed the failure paths.
What was built. Coordination systems putting multiple agents onto real development and business tasks, connecting APIs, data and AI steps into pipelines that run unattended, with explicit human review gates at the points where a wrong answer is expensive.
The outcome. Working automation, and a clear, tested view of where automation should not be trusted. That second part is what I bring to your operation.
About a week. I map where your operational time actually goes and what each hour costs you, then rank what is worth automating. You keep the map whether or not you hire me for anything after it.
We take the most expensive thing on that list and build it, onto the tools you already run. It goes live, your team uses it, and we measure the same number we measured in the audit.
Each workflow pays for the one after it. No long commitment up front, because you should not have to believe me before you have seen one of these actually work.
Nine years designing and building operational software, including six as founding designer on a cross-border eCommerce fulfilment platform, where I owned the order management, product research and merchant onboarding flows and built the design system the whole platform ran on.
I have also run my own direct-to-consumer store end to end, which means I have personally sat in the returns queue and the supplier chase. I know which parts of this genuinely hurt and which parts merely look inefficient on a diagram.
I design the system and I write the code, so there is no handoff between the person who understood the problem and the person who built the thing. You are talking to one person and that does not change later.
Based in India, working with merchants worldwide, async by default.
Everyone selling AI right now says yes to everything. Here is what I say no to, before you have paid me anything.