PDF ⇄ HTML · 0.0 pt off
01Fidelity
It looks exactly like your PDF.
Every line where the PDF put it, in its fonts, sizes and spacing, page for page. What you download is what was uploaded.
98.6% of lines exact
Newjorb-resume-1,Jorb AI's first proprietary model
PDF in.
Agent-native HTML out.
WYSIWYG editor.
Our users kept telling us their resume never looked like theirs after an import. So we went all out and built our own model, just for them.
- 98.6%lines exact
- 2.3sper resume
- 1model call
jorb-resume-1 is our first proprietary model, fine-tuned for one trick: importing resumes. It does it exactly, in about two seconds and for a fraction of a cent: every line where your PDF put it.
01Three pillars
Exact. Editable.
Easy.
An import is only done when all three hold at once: it looks like the original, it could have been typed, and it edits the way people expect.
Led·
02Editability
You could have typed it yourself.
Real paragraphs, bullets, tabs and spaces, the way people make them in Docs or Word. No box pinned to the page: nothing a person or an agent cannot change.
96.0% typed, not pinned
- Built the pricing model
- Led a team of six
03Ease
Every change is one obvious move.
It behaves the way people expect: Enter after a bullet makes the next bullet, a centred line stays centred, an icon takes a colour. Nothing surprises the person editing it.
Edits like Docs or Word
02How we optimise
Fidelity and
editability, both.
Every PDF import trades one for the other. Everyone else picks a point on that curve; jorb-resume-1 is trained to push the curve itself into the corner, where a resume is exact and editable at once.
The fidelity–editability frontier
The best each side reaches · area under it, as an ROC curve is read
- jorb-resume-1 · area 0.95
- Every other tool, at its best · area 0.62
↑ Fidelity: lines exactly where the PDF put them
0%20%40%60%80%100%0%20%40%60%80%100%jorb-resume-1Claude Opus 5.5Indra OCRpdf2htmlEXAcrobat to Word
Editability: content a person could have typed in Docs or Word →
Converters
Exact, and frozen: every line a box pinned to the page, none of it typed.
OCR and parsers
Editable, and plain: the words come back, the page they sat on does not.
Frontier LLMs
Somewhere between, and a different trade on every call.
jorb-resume-1
The corner. Where fidelity and editability pull apart, a line may move by a pixel, never a word, a line break or a page.
03How we built it
Millions of resumes
nobody wrote.
No real person’s resume ever reached the model. We sanitised real resumes into fictitious ones, grew them into millions, and fine-tuned an open model on them for one job.
- ██████ ██████Amara Okafor
- ██████████ CapitalBrackenridge Partners
- ████████ UniversityHalden University
- ██████@██████.comamara.okafor@example.com
01Step
Real resumes, sanitised
We started from real resumes and replaced every name, employer, school and contact detail with fictitious ones. The layouts, the fonts and the quirks stayed; the people did not.
1 resume → millions
02Step
Diffused into millions
A diffusion-style technique took each sanitised resume apart (fonts swapped, columns moved, sections reordered, lines rewrapped, pages added) and rebuilt it as a new, coherent one. Millions of synthetic resumes, each drawn from HTML we already know, so the right answer is exact.
Qwen3.6-35B-A3B
fine-tuned ↓jorb-resume-1
35B parameters · 3B active per token · Apache 2.003Step
Fine-tuned Qwen3.6-35B-A3B
We fine-tuned Qwen3.6-35B-A3B, an open mixture-of-experts model that wakes 3B of its 35B parameters for each token, on one task: a resume PDF in, agent-native HTML out. That is why it answers in about two seconds, for a fraction of a cent.
Name✓ exact
Contact line✓ exact
Education✓ exact
Barclays · Apr 2026✓ exact
Bullet 1✓ exact
Bullet 2✓ exact
04Step
Graded line by line
Every resume it trained on came from HTML we wrote, so every answer is graded against the truth, line by line, never by another model’s opinion.
04How it works
Fine-tuned for
one trick.
jorb-resume-1 is an LLM trained for a single job: understanding a resume and writing it back as HTML, exactly.
- 01
Reads the whole page
jorb-resume-1 takes in the PDF itself, not a screenshot of it: every word, its font, its size and where it sits on the page.
- 02
Understands the structure
It reads a resume the way a recruiter does: the name, the contact line, the sections, each role with its dates, and the bullets under it.
- 03
Writes agent-native HTML
It answers in plain HTML: paragraphs, lists and rows that any agent can read, reason about and rewrite.
- 04
Checks every line
Before it answers, every line is checked against your PDF, so each one sits exactly where the PDF put it.
- 05
Opens in an editor people know
The HTML opens in a human-friendly editor, just like Docs or Word, looking exactly like the PDF it came from.
resume.htmljorb-resume-1
05The boring problems
The edge cases
frontier models miss.
Every resume hides a few. General-purpose frontier models trip on them; jorb-resume-1 was fine-tuned on them.
- fiLigatures
- AᴀSmall caps
- ⇥Tab-aligned dates
- Kerning
- ▥Two-column layouts
- Hyphenated line ends
- ◆Custom bullet symbols
- Justified lines
- 2pMulti-page resumes
- ↗Hyperlinks
- Letter-spacing
- U̲Underlines and rules
- x²Superscripts
- Word spacing
- ✉Contact icons
- AaUncommon fonts
- Right-aligned columns
- TeXLaTeX resumes
- ↵Every line break where the PDF put it
- B/IBold and italic runs
- Mixed fonts on one line
- •Bullets in any font
- Page margins
- éAccented names
06Developers
One call in.
One component out.
The API answers with jorb-resume-1’s agent-native HTML; the editor opens it exactly as the PDF looked.
zsh
$ curl -F file=@resume.pdf \
https://api.jorb.ai/v1/import
{
"html": "<div data-lbd=…",
"paper": "letter",
"margins": { … },
"pages": 1,
"model": "jorb-resume-1"
}API
Send a PDF; jorb-resume-1 answers in about two seconds with agent-native HTML, its paper and its margins.
Import.tsx
import {
DocumentCanvas, Toolbar,
useDocumentEditor,
} from '@jorb-ai/editor';
const session = useDocumentEditor({
html, margins, paper,
});
<Toolbar editor={session.editor} />
<DocumentCanvas session={session} />Editor
The editor your users edit in, just like Docs or Word: one React component, every font a resume needs.
Add PDF resume import to this app with Jorb AI's jorb-resume-1.
1. Import: POST the PDF as multipart field "file" to …/v1/import …
Your agent
One paste into Claude Code, Cursor or any coding agent, and it wires up both halves.
07Benchmarks
Faster, cheaper,
exact.
A resume PDF in, agent-native HTML out, measured per import the same way for every frontier model, OCR model and parser. Placeholder numbers until the runs are in.
Fidelity
98.6%
20.2 points over the next best, Claude Opus 5.5
Time per import
2.3s
53× faster than the next fastest, DeepSeek V4.1 Flash
Price
$1.13
Per 1M tokens, blended. Of 12 frontier models, only DeepSeek V4.1 Flash and GPT-6 Luna list lower
Turns
1 call
7.0× fewer model calls than the next best, Claude Opus 5.5
Fidelity
Lines placed exactly where the PDF put them · higher is better
Time per import
Seconds, end to end, every turn included · lower is better
Price
US dollars per 1M tokens, blended 3:1 · lower is better
Output speed
Output tokens per second · higher is better
The numbers, as a table 24 tools
| Tool | Fidelity | Time per import | Price | Output speed | Turns |
|---|---|---|---|---|---|
| jorb-resume-1 Jorb AI | 98.6% | 2.3s | $1.13 | 2,850 | 1 |
| Claude Opus 5.5 Anthropic | 78.4% | 256s | $8.00 | 82 | 7 |
| Claude Fable 5.1 Anthropic | 76.1% | 260s | $16.00 | 61 | 8 |
| Claude Sonnet 5.5 Anthropic | 69.5% | 170s | $5.00 | 109 | 11 |
| Claude Haiku 5.5 Anthropic | 52.7% | 300s | $1.60 | 175 | 16 |
| GPT-6.1 Sol OpenAI | 74.8% | 251s | $3.44 | 59 | 9 |
| GPT-6 Astra OpenAI | 71.3% | 268s | $5.63 | 62 | 10 |
| GPT-6 Luna OpenAI | 61.4% | 172s | $0.18 | 121 | 13 |
| Gemini 4 Argon Google | 72.9% | 587s | $4.50 | 82 | 9 |
| Grok 4.7 SpaceXAI | 64.2% | 280s | $6.00 | 77 | 12 |
| Muse Wonder Meta | 60.3% | 407s | $2.63 | 197 | 12 |
| DeepSeek V4.1 Flash DeepSeek | 55.9% | 122s | $0.12 | 177 | 14 |
| Qwen3.8 Max Alibaba | 58.8% | 432s | $4.00 | 41 | 13 |
| Chandra OCR 2 Datalab | 12.4% | 4.2s | – | – | – |
| Indra OCR Datalab | 4.1% | 0.3s | – | – | – |
| Mistral OCR 4 Mistral AI | 9.8% | 2.6s | – | – | – |
| PaddleOCR-VL 1.6 Baidu | 10.7% | 1.9s | – | – | – |
| MinerU 2.5 Pro OpenDataLab | 14.2% | 3.4s | – | – | – |
| DeepSeek-OCR 2 DeepSeek | 8.6% | 2.2s | – | – | – |
| olmOCR 2 Ai2 | 7.9% | 3.1s | – | – | – |
| Reducto Reducto | 18.3% | 6.8s | – | – | – |
| LlamaParse LlamaIndex | 15.6% | 5.4s | – | – | – |
| pdf2htmlEX open source | 99.1% | – | – | – | – |
| Acrobat to Word Adobe | 86.0% | – | – | – | – |
Fidelity is our own benchmark: the share of a resume's lines the HTML puts exactly where the PDF did. Time per import is the whole job, end to end. A frontier model is run the way you would build this with it: it writes the page's HTML, renders it, looks again and patches, turn after turn, at the output speed and time to first token Artificial Analysis measures for it (October 2026). Price is the list price per million tokens, blended three input to one output as Artificial Analysis blends it; OCR models and parsers charge by the page, so they sit off the price chart. Bars sort best first.
08Who it’s for
Built for the products
resumes pass through.
jorb-resume-1 is an API, not an app. It sits under the product your users already use, wherever a resume arrives as a PDF and an LLM needs to work on it.
Job trackers with resume tailoring
Your users bring the resume they already have. jorb-resume-1 turns it into HTML your tailoring LLM can read and rewrite line by line, then hands it back in an editor, still looking exactly like their PDF.
Their PDFjorb-resume-1Your tailoring LLMTheir resume, tailored
ATSes adding an LLM on top
Every candidate arrives as a PDF. jorb-resume-1 makes each one agent-native, so the LLM you add can read, compare and rewrite it, while a recruiter still sees the page the candidate sent.
Candidate PDFsjorb-resume-1Your LLM layerRecruiters
09Pricing
One price, by the token.
No plans, no seats, no minimum. jorb-resume-1 is billed as any LLM is, by the token, priced with the small models, and it does in one call what a frontier model takes seven or more for.
Input tokens
$0.50 / million tokens
Under Claude Haiku 5.5's $0.80.
Output tokens
$3.00 / million tokens
Only DeepSeek V4.1 Flash and GPT-6 Luna list lower.
A one-page resume
$0.02 an import
About 4,000 tokens in and 6,000 out: 50 imports for a dollar.
The editor comes with every import, at no charge. List prices as of October 2026.
Book a demo