NLP, OCR and computer vision

Reading text, documents and images at scale. We extract data from invoices and forms, sort and score text, check images automatically, and send anything uncertain to a person.

What’s included

Document extraction and OCR

Invoices, purchase orders and forms turned into structured data.

Text classification and sentiment

Tickets, reviews and emails sorted and scored automatically.

Image recognition and visual checks

Products, defects or documents identified in photos.

Record matching

Duplicate customers and products found and merged across messy data.

Speech and Indian languages

Voice input and text in Hindi and other Indian languages.

When it makes sense

  • People retype data from PDFs and photos.
  • Thousands of reviews or tickets go unread.
  • The same customer exists several times in your systems.

What you get

  • Accuracy measured on your own samples
  • A review screen for low-confidence results
  • An API or batch job your systems call
  • Monitoring after launch

Tools we work with

We pick what fits your project and your team, and we’ll say why.

  • Python
  • PyTorch
  • Transformers
  • Wav2Vec2
  • OCR
  • Vision models

Related work

Built by our engineers before Orandic. None of these clients hired Orandic.

Machine learning

Native language from spoken English

Hears five seconds of English and predicts the speaker’s first language: Hindi, Kannada, Malayalam, Tamil or Telugu.

Harsh Sahu’s machine-learning project. The code and the experiment log are public.

Ramesh ko 500 rupaye udhaar

Here’s what I understood: Ramesh, udhaar ₹500. Confirm and save?

Voice app

DukaanMitra

A shop ledger for kirana stores. Say it in Hindi, Hinglish or English, “Ramesh ko 500 rupaye udhaar”, and it shows exactly what it understood before anything is saved.

Harsh Sahu’s project.

Questions about NLP and computer vision

How accurate is document extraction?

It depends on the quality and layout of your documents. We measure accuracy on a sample of your real documents before the build, and anything below the confidence you set goes to a person.

Does it work on handwriting and Indian languages?

Often, but not always reliably. We test on your own documents first and tell you what to expect before you commit.

Related services

Tell us what you need built.

A free 30-minute call with an engineer. You get the scope, timeline and price in writing before you commit, and a reply within one working day.