About
Jeff Combs, growth engineer.
Atlas Digital Design is a one-person technical department for small e-commerce businesses. Two live client engagements, everything below built by the person writing this, and a working record of about 29,000 words that I will share on request.
Definition
What a growth engineer is
A growth engineer is one person who builds and maintains the technical systems a small online business sells through, and who is accountable for what those systems earn. The role spans storefront code, product data, search, analytics and paid channels, on the principle that revenue faults usually sit between those areas rather than inside any one of them.
The title exists because the usual titles each describe half the job. A developer's remit is usually the build, and a growth marketer's is usually the number. Somebody has to hold both, or the questions that cross the boundary go unasked.
Comparison
How it differs from adjacent roles
| Role | Owns | Where the remit usually ends |
|---|---|---|
| Growth marketer | Channels, spend, messaging, the revenue number | At the theme code that decides whether an option renders |
| Front-end developer | The build, to the brief, to spec | At the brief. Whether the brief earns anything is usually someone else's question |
| Technical SEO consultant | Crawl, index, structured data, site health | At search. A feed defect or a broken conversion number sits outside it |
| Data or analytics engineer | Pipelines, models, the numbers being correct | Usually at read access to the thing generating the numbers |
| Growth engineer | All of the above, on one catalog, at once | Does not scale to a large team's throughput, and is not meant to |
Why one person
One system, and one person answerable for it.
I work across the storefront code, the product data, the measurement and the paid channels, and I stay responsible for how they work together. The faults that cost the most sit where two of those meet.
Sixty-six plan descriptions and thirty-seven sample-set descriptions rewritten in one voice in a single day, then every plan proofed field by field against thirty-six construction-set cover sheets, with fifty-three structured values corrected. A product feed rebuilt from scratch, validated against its specification with a strict parser, and cut over on a live sixty-six item catalog with zero warnings. Six years of revenue across five sales channels reconstructed from twenty-five thousand messages, with the July 2026 ledger matching the marketplace's own statement to the cent. Eight theme releases in September 2026 alone, each staged on an unpublished copy and verified by checksum.
Where the faults sat
The argument for one person is where the faults sat. A catalog this size was never going to hire five specialists anyway. The price defect in the shopping feed came from the way a sales channel app exports the catalog. Three dead features were one missing line in a layout file, found while building something unrelated. Years of the business's own revenue sat unread in an email archive, where no dashboard was going to look for it.
Each of those sits on a seam between two jobs. Opening the theme is outside a feed specialist's brief, and querying the order history to learn whether a shipped feature ever worked is outside an analytics contract. Split the work across four vendors and a defect on the seam has four places to sit unclaimed.
The agency that built a feed is rarely the one that discovers it has been advertising the wrong price. Somebody has to read the output from outside the tool, and that includes mine.
Auditing my own work
The same scrutiny, turned inward.
Some of what gets found here is in code I wrote. A system nobody audits drifts, and mine is no exception, so the method has to survive being pointed at its author.
- Every version of my revenue parser is diffed row by row against a frozen gold standard. That is what caught my own defect: bundled orders were having their base price dropped from every total, so six years of revenue read low across fifty orders, with a comment in my own source asserting the opposite behavior. The diff found it before anyone acted on the number.
- A gate I had written to show a $261.95 upgrade compared a product category against one exact spelling. Ten live pages spelled it differently, so the condition was false every time and drew nothing for fifteen days. Found by auditing the condition rather than looking at the page, because the page looked complete.
- Forty-five of my own pages carried a crawl block and a noindex tag at once. A noindex only works on a page the crawler is allowed to fetch, so the block guaranteed the tag was never read and the URLs were indexed anyway, title-less. Two safeguards I had put there myself, canceling each other out.
- A comment in my own code read "verified: the label is a static placeholder, not a floating one." It floats. The line that comment justified was the defect it was hiding, and both came out.
Engagements
Two live clients. Both checkable.
Intertwine House Plans
Sole growth engineer for a 66-plan Shopify catalog of architectural house plans, competing for the same searches as the national aggregators that list the same plans. Storefront, product feeds, structured data, paid search, lifecycle email, digital fulfillment and the revenue ledger underneath all of it. Since July 2024, continuous from May 2026.
| Product feed | Rebuilt from scratch. 66 of 66 items ingested, zero warnings, validated by strict parser and per-field assertion |
| Revenue data | Six years and five channels rebuilt from 25,562 messages of primary source, with the July 2026 ledger matching the marketplace's own statement to the cent |
| Layout stability | Cumulative layout shift on the modification form taken from 0.463 to 0.000 |
Bethel Lutheran Church, Roseville
Technical remediation on a nineteen-page site running a legacy stack, taken on ahead of a proposed rebuild. Ownership of the domain was traced and separated from the hosting account before anything was touched, a site-wide duplicate-content fault was corrected across every page, and titles and descriptions were written from each page's own words rather than invented. Since August 2026. Remediation in progress, rebuild scoped and not started.
Software
Built from nothing. Three in production, one waiting.
| What | Size | What it does |
|---|---|---|
| Revenue ledger | 8,250 lines Python, four major versions | Reads six years of order email across five channels under each channel's own royalty rules, every version diffed against a frozen gold standard |
| Split Engine | 2,967 lines Python, white-label | Revenue splits allocated from gross in one largest-remainder pass so the parts reconcile exactly. A weight set that does not total 1.0 stops the app rather than mispaying someone |
| Study-set pipeline | PyMuPDF and PIL, nine products | Reduces 36×24 construction sets to 11×17 samples, driven by per-product JSON so a new product is a data file and never a fork |
| Pin Board | Native macOS, OAuth 2.0 | Plans and inspects a 5,690-pin account with no destructive API call available to it. Publishing is built and has not been switched on |
The architecture, the test strategies and the full inventory of languages, platforms and instruments are on the engineering page.
Method
The rules I work by.
- Verified and assumed are distinguished in writing, and which is which is marked.
- A fix does not ship until it has been tested by the mechanism that would catch its failure mode. A guard that never matches reports nothing at all.
- Every change is proposed before it is made and logged with its rollback date at the moment it ships, so a month of work can be unwound by reading one table.
- Adversarial review is sought and acted on, including when it invalidates the work. An outside audit's genuine hits get credited as prominently as its misses get corrected.
- A diagnosis that the instrumentation contradicts is retracted in the same exchange it was offered, and the record is edited to say so.
- No claim goes on a client's site that the business cannot stand behind. That rules out "instant" for something a person delivers by hand, and any credential the owner does not hold.
Questions
Asked often enough to answer here.
What a growth engineer does day to day
Reads systems that are already running, finds where they are losing money silently, fixes those, and builds what is missing. On a given week that can mean writing Liquid, validating a product feed against its specification, querying six years of orders to settle an argument, rewriting ad copy, or shipping a Python application because a spreadsheet stopped being enough.
How this differs from SEO
Search is one of eight disciplines here. Most of the faults I find have never been near a search console.
Working alongside an agency you already have
It works on the right split. I take the wheel on the digital architecture, which is the storefront data, the feeds, search, tracking and the advertising, because the value is in holding those keys at once. A front end developer you already trust is no obstacle to that. A full rebuild including the design is the case where I need the reins.
The size of business this suits
Small catalogs selling their own product, where one person can hold all of it and the margin is good enough that a silent fault is expensive. If you run a thousand SKUs with an in-house team, the audit is better done by that team.
Why there is no price on the site
Because there is no fixed job. A nine-product catalog on a clean theme and a six-hundred product catalog with three vendors writing to it are not the same week of work, and quoting them identically would mean overcharging one of them. I look first, then quote a single figure.
The shape of the arrangement is fixed even though the number is not, and it is written out in full on the teardown page. In short: the teardown is quoted per site and credited against your first month. The rebuild runs on a retainer for around six months. After that the fee drops and a share of the growth above your existing baseline takes its place, with the baseline never taxed.
What the growth share is calculated on
Only on the part that was not there before. Whatever the business was already earning when I arrived is the baseline, and it is never taxed. If I do not move the number I am paid the monthly fee and nothing else, so I have no reason to make the work take longer.
What happens when I am wrong
It gets retracted in writing, in the same document, with the correction beside the original. Four examples are listed above and two more are on the home page.
Who owns the work
You do: the code, the documents and the data. Nothing is built on a platform only I can operate, the change log lives where you can read it, and if the engagement ends you keep a working system.