About Novycom
Novycom grew out of running cross-border commerce, not out of a research group. We spent years inside marketplaces, 3PLs and supplier spreadsheets before we built anything with a model in it — which is why our first instinct is to ask what the process actually looks like, rather than which framework to use.
How we operate
These aren't values on a wall — each one has cost us a project at least once, and we kept them anyway because the alternative is a client who stops trusting the system.
01
If a process is low volume, has no verifiable right answer, or fails expensively, we say so on the first call. A vendor who accepts every brief is telling you something about their pipeline, not your problem.
02
Agreed in advance, in plain terms, against a test set from your own history. If the pilot misses it, we tell you rather than adjusting the definition of success.
03
Source, prompts, eval sets, documentation, infrastructure definitions. No licensed runtime, no proprietary layer that makes leaving expensive. We'd rather be kept than locked in.
04
Postgres before a new vector database, a scheduled job before an agent, a script before a model. The interesting choice is usually the expensive one to maintain.
05
When your team can run it, we write the runbook and step back. Retainers that outlive their usefulness are how agencies decay.
06
Monthly, including the months where accuracy dipped or cost per task rose. Reporting only the good months makes the good months meaningless.
The team
Eighteen people in Kuala Lumpur. No offshore delivery layer, no account manager between you and the people writing the code — you talk to whoever built the thing you're asking about.
Backend and data engineers who have run production systems before, not prompt specialists. Median eight years' experience, mostly from commerce and logistics platforms.
People who have worked marketplace back offices themselves. They run the scoping week, because knowing which questions to ask matters more than knowing the answers.
Native writers for Chinese, Bahasa, Thai and Vietnamese. Customer-facing output is written per market, never machine-translated from English.
Who you work with
No account layer. The person who scopes your project is in the room when it's built, and the engineer who built it is on the call when you have a question about it.
Founder & principal engineer
Twelve years across marketplace operations and backend engineering. Runs scoping personally on every engagement, which is why we turn work away.
Head of delivery
Former warehouse systems lead at a regional 3PL. Owns the accuracy bar on every project and has stopped more pilots than he has shipped.
Data engineering
Entity resolution and pipelines. The reason projects here start with catalogue work instead of discovering the problem in week six.
Language & content
Leads per-market writing across Chinese, Bahasa, Thai and Vietnamese. Vetoes anything that reads like it came out of a translation engine.
Where we differ
Not a comparison table — just the differences clients tell us they felt in the first month.
You pay a fixed fee and receive a document you can take to another vendor. That changes the incentive from winning the project to being right about it.
Shadow mode against a written bar means "it didn't work" is an available outcome. Most vendors structure pilots so that outcome can't happen.
Building and running are different jobs with different costs. Bundling them hides the ongoing number until you're committed.
How we got here
We started as the internal team that had to make this work for one business, which is why the firm is shaped the way it is.
2022
Where this began
Three of us were running back-office operations for a cross-border retailer. Order exceptions across five marketplaces were consuming most of a team's week, and every tool we evaluated assumed a process we did not have. We built our own because there was nothing to buy.
2023
By accident
Two companies in the same warehouse park asked whether the thing we had built could be pointed at their queue. It could, but only after we rewrote most of it — which taught us the durable lesson that the integration layer is the product and the model is a component.
2024
Two failed projects
We took on two engagements we should have declined: one where nobody could define a correct answer, and one where the volume never justified the build. Both were delivered, neither was used. The scoping week exists because of those two, and it has been fixed-fee ever since.
2025
The turning point
We started shipping the evaluation suite as a deliverable rather than an internal tool. Renewals went up, arguments about whether it was working went away, and two clients used the suite to hold us to a number we had missed. That was uncomfortable and correct.
2026
Where we are
Eleven people, working on six to eight engagements a year. We have turned down growth that would have required hiring faster than we can train, because the failure mode of this business is a team that ships things it does not understand.
We have turned down growth that would have meant hiring faster than we can train. The failure mode of this business is a team shipping things it does not understand.
Working agreements
These are in our contracts, not just on this page. If we break one, say so and we will fix it or refund the sprint.
If a build is behind, an accuracy bar will be missed, or an approach is not working, you hear it in the week it becomes clear — not at the milestone. We have absorbed the cost of a missed fortnight rather than move a deadline quietly, and we would rather do that than manage your expectations.
Your code, your repository, your cloud account, your model API keys. No proprietary runtime, no platform fee, no component that stops working if you stop paying us. Cancel the retainer and the system keeps running.
The people on the scoping call are the people who build it. There is no separate delivery team, no offshore hand-off, and no account manager relaying questions between you and someone you have never met.
Including from us. About a third of what we are asked about does not need a model, and roughly one in five clients takes the scope and builds it internally. Both are stated outcomes of the scoping week, not concessions extracted later.
Accuracy is reported on a random sample we did not tune against, re-drawn quarterly. Cost per task includes human review time. If a number moves the wrong way, it appears in the monthly report with an explanation before you ask.
Handover material is written during the build, not at the end, and the runbook is tested by someone who did not write it. A client leaving to run the system themselves is the outcome we designed for.
The team
Eleven people in Kuala Lumpur. The hiring bar is unusual in one specific way, and it shows up in the work.
Seven of eleven have worked in a warehouse, a customer service floor or a finance team. It is why our first question is what the process actually looks like at 4pm on a Friday rather than what the API returns.
Every hire writes a one-page technical explanation as part of the process. Runbooks, decision notes and monthly reports are deliverables, and an engineer who cannot explain a trade-off in writing cannot hand a system over.
We ask candidates to describe a project they argued against. Someone who has never pushed back on a bad brief will not do it on yours, and the most valuable thing we sell is an honest opinion in week one.
We are not hiring at the moment. When we are, it is posted here rather than through recruiters, and the process is four conversations with the people you would actually work beside.
Where we are
One office, one time zone, and a deliberate limit on how far we work from it.
The office is in Kuala Lumpur, and for most clients that is close enough to stand in the warehouse during the scoping week. Being able to watch the process at four on a Friday afternoon is worth more than any workshop.
It is also why we cap how far we work from here. Past about four hours of overlap the coordination cost stops being fair to the client, and we would rather say so than take the engagement.
On-site for the scoping week as standard, and we will come back when something is going wrong. Most of our work is here, and being able to stand in the warehouse is a large part of why it works.
Regular travel for scoping and go-live, remote for the build. We work in English and Mandarin directly, and bring in native reviewers for Bahasa, Thai and Vietnamese content rather than trusting a model's fluency.
We take remote engagements where the overlap is at least four hours and the client has a technical contact in-house. Beyond that the coordination cost stops being fair to you, and we will say so rather than take the work.
Work with us
Thirty minutes, no NDA required, and an honest answer about whether we're the right people for what you have in mind.