Skip to content
TPA Ventures

Applied AI

ML Engineer, Document Understanding

Extract scores, findings and recommendations from psychoeducational evaluations and assessment reports with source-level traceability, so every baseline in a draft can be verified.

Team
Applied AI
Location
Mumbai, India (Fort, hybrid)
Type
Full-time

What you will do

  • 01Build extraction pipelines for scanned and digital evaluation reports, including tables of standard scores and percentiles
  • 02Design evaluation sets and metrics for extraction accuracy, with zero tolerance for invented values
  • 03Ground language model drafting in extracted data with citations back to the source page
  • 04Work with security engineering so no student data is used for model training
  • 05Partner with content specialists to review edge cases from real assessment formats

What we are looking for

  • 014+ years in applied machine learning or NLP, with production document AI or OCR experience
  • 02Strong Python and experience evaluating LLM systems rigorously
  • 03Experience with layout-aware models, table extraction or information extraction
  • 04Judgment about precision versus recall in high-stakes settings
  • 05Familiarity with assessments such as the WISC-V, WJ IV or CELF-5 is a plus

How to apply

Email the hiring team. A person reads every application within five business days and replies either way.

Apply for this role

Sends to careers@tpaventures.si with the subject “Application: ML Engineer, Document Understanding”.

What to include

  • A resume or a link to your profile. No cover letter required.
  • Three to five sentences on one thing you built, taught, shipped or fixed that you are proud of, and what you would do differently now.
  • Where you are based and the hours you prefer to work, so we can plan the overlap window.
  • For engineering and design roles, a link to code, a portfolio or a write-up. For content roles, the state whose rules you know best.

Other open roles

Want to see the product before you apply?

Watch a draft built from a real evaluation format, with every baseline cited, in a twenty-minute walkthrough.