Can Glimpse review your CV layout from an image?

What exactly did Elorian launch, and can it review a CV?

Glimpse can tell you how your CV looks. It can't tell you how your CV parses. Elorian, a Palo Alto startup that raised a $55 million seed round at a $300 million valuation before shipping a product, released the model on 16 October 2026 as a research preview — API access now, chat interface by waitlist. The pitch is genuinely different from the usual multimodal setup. Most models glance at an image once, flatten it into tokens, then reason in text, which Elorian's CEO compares to describing a photo from memory. Glimpse keeps going back to the source mid-reasoning: cropping into a region, pulling a specific video frame, running code to measure or compare what it's looking at. The company calls that a visual scratch space, and it's the architecture, not a bolt-on.

For a CV, that mechanism matters more than it sounds. A CV is a layout problem as much as a writing problem — two columns, a sidebar, a skills bar chart, uneven margins, a second page carrying four lonely lines. A model that can crop into your header and actually measure the gutter between columns is doing something a text-only assistant cannot do at all. Elorian's own evaluations put Glimpse ahead of GPT-6 Astra on scientific chart reasoning and short-form video captioning, with results it describes as competitive on other visual reasoning tasks. Those are internal numbers rather than an independent lab's, and the model is substantially smaller than the frontier systems it was compared against, trained by a small team on a few hundred GPUs.

Here's the catch, and it's the point of this whole article. Glimpse was built to close the gap between what a picture contains and what a model understands — the founding anecdote is photographing a finished board game, asking a frontier model who was winning, and getting nothing useful back, because every game carries its own iconography. Your CV has the opposite problem. The picture is fine; the file is broken. A vision model sees the document a human sees, not the text an applicant tracking system extracts, and in CV land those are routinely two different documents. We reached a similar verdict testing an offline model running on Apple silicon: a general-purpose model can critique writing and layout, and it still can't run the check that decides whether you're screened out.

What can an image-reasoning model actually see in your CV?

An image-reasoning model sees your CV the way a recruiter's eye sees it in the first few seconds: the shape of the page, where the heavy blocks of text sit, and whether anything stops the eye on its way down. That's not a small thing. Crop-and-measure reasoning means a model like Glimpse can go past vague praise for "good hierarchy" and get specific — the gap between your two columns is too tight, your section headings are only a point larger than body text, your margins shrink between page one and page two, your bullets run to four lines each, and that second page is two-thirds empty. Specific feedback is the only kind you can act on.

It also reads the parts of a CV that are pictures rather than words. Skill bars, five-dot proficiency ratings, a donut chart of "time spent per tool", timeline graphics, icons standing in for "phone" and "email", a headshot — a text-only assistant sees none of it, so it cheerfully praises a CV it never fully read. A vision model sees all of it, and it'll tell you that four dots next to Python says nothing a hiring manager can use. It's good at consistency too: three different bullet glyphs, two font families fighting each other, dates that switch from "Jan 2022" to "01/2022" halfway down the page, a job title bolded on one entry and plain on the next.

An opinion you can hold me to: for pure visual critique, a strong image model now beats most of the free feedback you'll get from a forum thread or a kind friend. It's patient, it's specific, and it has no taste to defend. Ask it what you read first, second and third, and the answer tells you whether your layout is pulling its weight. Then pair it with a structural pass, because design feedback and screening feedback are different jobs — a twelve-point CV self-check walks through the recruiter's seven-second scan and the red flags that get a CV binned outright. Between the two, you've covered the human reader properly.

What does a vision model miss that an ATS check catches?

What Glimpse misses is the file. A vision model reasons about a rendered picture; an applicant tracking system reads the text layer underneath it, and those two readings disagree far more often than template designers admit. A two-column CV looks balanced on screen and can extract as interleaved nonsense, with your job title landing mid-sentence inside a skills list. Text placed inside a shape or text box renders perfectly and extracts as nothing at all. Contact details tucked into a header or footer get dropped by plenty of parsers. A skills table collapses into one unreadable line. A custom font can produce garbled characters on copy. None of that moves a single pixel, which is exactly why the screenshot still looks great.

Keywords are the second blind spot, and it's the one that costs interviews. A screenshot critique can't know that the job ad says Kubernetes while your CV says K8s, that the posted title is "Data Analyst" and you called yourself an "Insights Specialist", or that the ad names a certification you actually hold but never listed. That's a text-against-text comparison with one specific job description, not a judgement about how a page looks. Hebrew and bilingual CVs bring a third blind spot. Hebrew that reverses character order on extraction looks completely normal as an image: the model reads the pixels and reports clean, well-set Hebrew, while the employer's system receives scrambled text.

Closing that gap is the job an ATS-aware check exists to do, and it's why HRLens asks for your file rather than a picture of it. Upload a PDF, a Word document or an image, and the free analysis returns an overall score out of 100, five category scores — experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility — plus a visual layout analysis. You get the eye test and the file test from one upload, which is the combination a screenshot can never give you. If your CV came out of a design tool full of columns and icons, start with why a beautiful CV scans badly before you rewrite a single bullet.

Should you upload a screenshot or the actual CV file?

Upload the real file, every single time. Use an image only when no file exists — a printed CV from a previous decade, a document someone else controls, a photo taken at the kitchen table. A screenshot discards precisely what automated screening reads: the text layer, the reading order, the embedded fonts, the character encoding. You hand over the picture and throw away the evidence. It's the equivalent of asking a mechanic to diagnose your engine from a photo of the hood. The picture tells you something honest about the paintwork; it tells you nothing about why the car won't start. Same CV, two completely different diagnoses.

When you do upload an image, optical character recognition rebuilds the text from the pixels — HRLens supports image uploads, Hebrew OCR included — so you still get a real reading of what the page says and how it's written. What you can't recover is the original file's parsing behaviour, because OCR has already rewritten the text for you. Which brings up the trap worth naming loudly: never send an employer an image-based CV. If your "PDF" is a scan, a phone photo, or a flattened export from a design app, there's no text layer to extract and some parsers hand the recruiter an empty record. A beautiful CV that scans as a blank page is a CV nobody reads.

Israeli job seekers hit this harder than most. A bilingual CV — Hebrew body text with English tool names, company names and job titles — mixes right-to-left and left-to-right runs inside single lines, and that's exactly where extraction breaks: numbers flip, brackets face the wrong way, a date reads backwards, a phone number arrives reversed. Employers here screen through systems like Comeet while candidates apply through AllJobs, Drushim and LinkedIn, so your file meets a parser long before a human opens it. A vision model looking at a screenshot of that CV will tell you the Hebrew looks clean, and on screen it genuinely is. The damage happens on the way in.

How should you combine a visual check with an ATS check?

Use them in this order: fix the file first, polish the picture second. Run the actual document through a parser-aware check before you touch the design, because every structural fix changes the layout anyway — collapsing two columns into one, lifting your phone number out of the header, swapping an icon for the word "Email", turning a skills table into plain lines. Polishing the visual design of a CV that doesn't extract is sanding a door that doesn't close. Once the structure is sound and the text layer is clean, a vision model earns its place as the last reader before you hit send. That order also saves you work: you redesign once instead of twice.

In practice it's four steps. First, run your CV through a free analysis and read the ATS compatibility score before you look at anything else. Second, paste the job ad as text — the text, not the link, because LinkedIn and Indeed block fetching — so the missing skills come back tied to the role you're actually chasing. Third, fix the structure: one column, no text boxes, contact details in the body, real section headings, dates in a single format. Every analysis also generates ATS-friendly rewritten CVs across six templates that mirror the job's keywords, which saves you the rebuild. Fourth, screenshot the finished page and ask an image model what it reads first, second and third.

Keep Glimpse's status in mind before you build a job search around it. It shipped as a research preview: API access at $0.60 per million input tokens and $3 per million output tokens, chat interface still waitlisted, and the benchmark wins are Elorian's own figures rather than an independent lab's. Its stated targets are robotics, engineering, medicine and scientific analysis — not hiring. Beating a frontier model on scientific chart reasoning is a real achievement, and it tells you nothing about whether a recruiter's system will read your employment dates in the right order. Treat a vision model as a design critic with unusually good eyes. Treat the file check as the thing that decides whether anyone sees the design at all.

Frequently asked questions