Tested on : https://dsel.education.gov.in/careers/index.html -> vol 2 pdf (page 23-120)

This project converts a scanned career guidance PDF into structured JSON data that can be used by a React UI or any other application.

The current pipeline is:

Scanned PDF -> Page Images -> PaddleOCR -> Column-aware Parser -> careers_data.json

• Reads a scanned PDF career book.
• Converts every page into an image.
• Runs OCR using PaddleOCR.
• Uses OCR box coordinates to reduce multi-column text mixing.
• Extracts structured career data into JSON.
• Generates fields such as title, description, personality traits, education path, salary, scholarships, work environment, growth path, institutes, and example profile.

The latest run generated:

Total careers:         39
Descriptions:          39
Personality traits:    28
Education pathways:    38
Degrees:               39
Entrance exams:        39
Salary ranges:         37
Scholarships:          39
Work places:           36
Work descriptions:     38
Growth paths:          39
Government institutes: 39
Private institutes:    39

Some values still depend on OCR quality and page layout. The script keeps a raw_text field for every career so missing or imperfect structured fields can still be recovered.

This project was tested on Windows with:

• Python 3.12
• PaddleOCR 3.5
• PaddlePaddle GPU 3.2.0
• NVIDIA RTX 3050 4 GB
• CUDA 12.6 runtime from Paddle wheel
• Poppler for PDF-to-image conversion
• uv for creating the working Python environment

The original env folder used Python 3.14, but PaddlePaddle did not provide compatible wheels for Python 3.14. The working setup uses ocr-gpu-env with Python 3.12.

Open PowerShell in the project folder:

cd C:\Users\stiwa\Downloads\Roadmap

Run the extractor:

$env:PADDLE_PDX_DISABLE_MODEL_SOURCE_CHECK='True'
@('Enginnering.pdf', '') | .\ocr-gpu-env\Scripts\python.exe ocr_pdf_to_json_extractor.py

The second empty input uses the default PaddleOCR language, en.

After the script finishes, it writes:

careers_data.json

On the RTX 3050 4 GB setup, the full 98-page PDF run takes around 4-6 minutes.

The script intentionally runs OCR sequentially instead of using multiple OCR threads. This avoids GPU memory crashes on 4 GB VRAM cards.

Check that JSON was generated:

Get-Item careers_data.json

Check field coverage:

$data = Get-Content careers_data.json -Raw | ConvertFrom-Json
$careers = @($data.careers)
"Total: $($careers.Count)"
"Descriptions: $(@($careers | Where-Object { $_.description }).Count)"
"Traits: $(@($careers | Where-Object { $_.personality_traits.Count -gt 0 }).Count)"
"Education: $(@($careers | Where-Object { $_.education.pathways.Count -gt 0 }).Count)"
"Salary: $(@($careers | Where-Object { $_.salary_range.min -ne $null }).Count)"
"Institutes: $(@($careers | Where-Object { $_.institutes.government.Count -gt 0 }).Count)"

CareerGuidanceUI.jsx expects career data in the same shape as careers_data.json.

In a React project, copy:

• CareerGuidanceUI.jsx
• careers_data.json

Then import and render the component according to your app structure.

• Keep the PDF in portrait orientation for best extraction.
• Multi-column pages are handled by coordinate-aware OCR sorting, but complex page layouts can still create noisy fields.
• raw_text is intentionally preserved for every career so the data can be re-parsed later without rerunning OCR.