My Résumé Is a Lossy Compression Algorithm

Audio article

Uses the speech voice supplied by your browser or device.

The last time I tried to compress myself into two pages, I lost approximately forty percent of my actual mass. Not metaphorically. I mean the document itself sat there like a deflated balloon, all the air punched out, and what remained was technically legible and fundamentally dishonest. Fifteen years of American healthcare research — hospitals, databases, MUMPS, SQL, interoperability standards that only make sense if you have ever watched one hospital system attempt to communicate with another hospital system and fail in ways that would make the Tower of Babel look like a well-run conference call — all of it squeezed into bullet points so terse they could have been written by a telegram operator who charges by the word and holds a personal grudge against you.

A résumé is not a biography. Let us get that out of the way before anyone starts feeling warm and fuzzy. A résumé is a compression format. And like every compression format, it makes decisions about what to keep and what to throw away, except here the encoder is a terrified human being guessing what another terrified human being, or increasingly a machine, might consider relevant.

Continuity survives beautifully. Interruption looks suspicious. Familiar job titles — oh, those precious, anodyne, soul-murdering job titles — compress like JPEGs of sunsets. Everybody knows roughly what they are supposed to mean. But hybrid knowledge? Migration? A failed business? Five technologies learned because somebody had to learn them? A medical catastrophe? A decade that makes perfect causal sense when narrated aloud and looks like a crater when represented as two dates separated by white space?

Into the bit bucket.

MUMPS is a particularly good example. The language was born at Massachusetts General Hospital in the 1960s, when computers still looked like refrigerators designed by military contractors, and descendants of that ecosystem went on to underpin major healthcare information systems for decades. You can therefore have the peculiar experience of maintaining software whose intellectual ancestry predates the moon landing while being told by somebody with twelve months of recruiting experience that your skills appear “dated.”

The résumé has no field for irony.

I spent years commuting between Austin and San Antonio. That is a sentence that compresses into nothing. What it actually meant was that I lived inside a car for what felt like geological epochs, watching the Texas landscape scroll past like a screensaver designed by someone who had never experienced joy, and somewhere in that liminal asphalt purgatory I learned things about healthcare data that no classroom would ever teach me.

But try putting that on a résumé.

“Commuted extensively.”

Excellent.

So did every other poor bastard who ever held a job in Texas.

The résumé does not care what you learned in the car. It cares that you were employed, that the dates line up, that the verbs are vigorous, and that at no point did your life develop the insolence to become nonlinear. If your chronology resembles a ruler, congratulations. If it resembles an ECG, somebody in Human Resources would like an explanation.

And now — now! — we have decided that the solution to this grotesque oversimplification is to insert artificial intelligence into the hiring pipeline.

Because if there is one thing guaranteed to restore information discarded during compression, it is another layer of computation operating on the compressed output.

This is like photocopying a photocopy of a photocopy and then asking a very confident robot to reconstruct the face of the original person. Perhaps it can infer something useful. Perhaps it can identify structure a human skim would miss. Perhaps it can compare skills more consistently than an exhausted recruiter at 4:47 on a Friday afternoon.

But it cannot recover information that was never supplied.

That part matters.

The input is already mutilated. The résumé has already performed its little administrative vivisection. It has stripped away context, narrative, causality, accident, illness, migration, obsession, failure, reinvention, and the hundred small reasons one year followed another. What remains is a skeleton of dates and nouns and verbs, dressed in Calibri and instructed to look employable.

Then we ask a machine to evaluate the skeleton and tell us what sort of animal it used to be.

A ten-year career gap might contain almost anything: illness, caregiving, migration, financial collapse, self-directed study, family obligations, freelancing, a business that never became a LinkedIn announcement, or simply a prolonged period in which life behaved like life instead of a project plan.

A responsible machine should not look at the blank space and hallucinate which one occurred.

That would not be understanding. That would be fortune-telling with GPUs.

What it could do — what would actually be useful — is preserve the uncertainty. Ask for context. Allow explanation. Recognize that “absence of conventional employment” is not synonymous with “absence of activity,” and that chronology contains less information than recruiters have collectively agreed to pretend it contains.

Likewise, an old technology is not automatically an obsolete mind. Somebody who maintained MUMPS-based healthcare systems may possess an unusually deep understanding of persistent clinical data, transactional workflows, hierarchical storage, legacy integration, and the peculiar reality that hospitals do not get to rewrite civilization every eighteen months because Hacker News has become bored with the current JavaScript framework.

You cannot infer all that merely from the word “MUMPS.”

But you cannot infer incompetence from it either.

The résumé invites both mistakes.

And then there is the famous recruiter skim.

One much-cited 2018 eye-tracking study from the career site Ladders put the average initial résumé screen at 7.4 seconds. Seven point four. Not a universal constant of nature, not Avogadro’s number for Human Resources, not proof that every recruiter on Earth operates with the attention span of a startled squirrel — but a wonderfully bleak measurement nonetheless.

Seven point four seconds.

That is barely enough time to regret opening the PDF.

Seven point four seconds to communicate the entirety of your professional existence to somebody who may be looking primarily for recognizable anchors: current title, employer, previous title, dates, education, familiar skills. Your fifteen-year education in how institutions actually malfunction has roughly the same opportunity to introduce itself as a YouTube preroll advertisement.

And if the human being does not inspect you first, software may.

“Applicant tracking system” is itself a broad and slightly deceptive term. Some systems are mostly filing cabinets with ambitions. Others can parse résumés, search them, apply knockout questions, enforce employer-defined requirements, score candidates, or help rank applicants. Increasingly, algorithmic and AI-assisted tools can participate elsewhere in recruiting and screening as well.

Which is quite different from the popular mythology in which a malevolent ATS reads “structured query language,” fails to find the sacred letters “SQL,” and personally launches your résumé into the sun.

The machinery is more complicated than that.

Unfortunately, complicated machinery can still produce stupid outcomes.

A parser can misunderstand formatting. An employer can configure crude filters. A screening criterion can be a bad proxy for the thing the employer actually wants. Historical hiring data can contain historical preferences, and historical preferences are merely old prejudices wearing business casual. Automation does not create wisdom by multiplying a bad assumption by ten thousand applications per hour.

It creates throughput.

Somewhere a product manager will call this “efficiency.”

But here is the question that keeps me up at night, the one that slithers into my skull and coils itself around my prefrontal cortex like an unwelcome houseguest who has decided to redecorate:

Could AI do the opposite?

Not magically reconstruct private facts from blank spaces. I do not want a hiring model looking at an employment gap and deciding, through some statistical séance, that I was depressed, caring for a parent, alcoholic, incarcerated, meditating in Bhutan, or breeding ornamental shrimp.

That is not empathy.

That is surveillance wearing empathy’s trousers.

But could a system be built to understand that missing information is missing information?

Could it distinguish “I do not know why this gap exists” from “therefore this gap is bad”?

Could it recognize skill adjacency instead of exact vocabulary? Could it notice that somebody who has crossed several generations of healthcare technology may possess transferable architecture knowledge even when the nouns on the résumé no longer match the nouns in the job advertisement? Could it let a candidate explain discontinuity without forcing the explanation into that ghastly little box labeled “Additional Information”?

Could it understand that an immigrant career is often not broken but translated?

Translation always loses something.

Sometimes there is no equivalent word. Sometimes an American employer does not recognize the institution, title, technology, hierarchy, credential, or cultural signal that would have been instantly legible somewhere else. Sometimes a senior person becomes junior simply by crossing an ocean. Sometimes competence survives intact while its labels are confiscated at customs.

A résumé records the new label.

It does not record the customs officer.

The word “compression” comes through Latin from comprimere: to press together.

That is exactly what the résumé does. It presses a life together until the life is flat enough to slide through the mail slot of corporate attention.

And the tragedy is not that compression happens. Compression is necessary. Nobody wants my forty-seven-page professional Bildungsroman complete with appendices, childhood influences, database schemas, psychiatric weather reports, and a graph of caffeine consumption against SSIS package failures.

The tragedy is that we forget compression happened.

We begin treating the compressed representation as though it were the thing itself.

The résumé says “gap,” therefore nothing happened.

The résumé says “legacy technology,” therefore obsolete.

The résumé says “consultant,” therefore perhaps unemployed in a blazer.

The résumé says “ten years,” but cannot tell you whether those ten years contain repetition or depth.

The résumé says “SQL,” but cannot distinguish between somebody who completed an online course last Tuesday and somebody who has spent years elbow-deep in production data trying to determine why 187,000 patient records have suddenly acquired the same discharge date.

Same token.

Different information.

That is lossy compression.

And now we are building increasingly sophisticated systems atop it.

I am not optimistic. I want to be. There is genuinely a version of AI-assisted hiring that could be less stupid than what preceded it. Machines can search large pools consistently. They can map related skills. They can surface candidates a recruiter might never manually reach. They can potentially make explicit the criteria by which candidates are being compared rather than leaving everything to the mystical vibrations of “culture fit.”

But there is another version.

That version industrializes the existing assumptions.

Feed it historical outcomes and it learns which kinds of trajectories historically received approval. Give it crude proxies and it optimizes the proxies. Tell it that uninterrupted employment, prestigious employers, recognizable titles, fashionable technologies, and conventional progression correlate with previous hiring decisions, and the machine does not need to hate unusual people.

Hatred would require far too much personality.

It merely needs an objective function.

A mind that does not fit the template — whether because of ADHD, bipolar illness, temperament, circumstance, obsession, migration, illness, poverty, caregiving, or simply a lifelong allergy to proceeding in the approved order — can produce a career that looks less like a ladder than a fractal.

Intricate.

Recursive.

Possibly beautiful.

Absolutely infuriating to summarize in six bullet points.

The hiring machine looks at the fractal and may see noise because noise is what information looks like when the decoder does not possess the codebook.

It looks at a life lived in depth and sees a gap.

It looks at a person compressed by migration, illness, circumstance, fashion, economics, and the simple brute fact of remaining alive while industries rename themselves every six years, and it asks whether the formatting is ATS-friendly.

Then, perhaps, comes the email.

“We regret to inform you.”

I regret to inform the machine that the problem may not be entirely on my end.

A résumé is an approximation of a person.

A useful approximation, sometimes.

A necessary approximation, probably.

But still an approximation.

The danger begins when the machine forgets this.

The greater danger begins when we do.

---

P.S. Claude Shannon’s 1948 work gave information theory its mathematical spine: for a source with a given entropy, there is a lower bound on the average number of bits required for lossless coding under the usual assumptions. A résumé is not literally a Shannon code. Human lives do not arrive as stationary information sources, and Human Resources does not possess the decoder.

So no, every résumé is not an attempt to “violate Shannon’s limit.”

The metaphor is better than that.

If you insist on representing a high-dimensional human life in two pages, you must decide what information to discard.

Then somebody else decides whether the missing information mattered.

And now we are teaching machines to help.