Bachelor's or master's thesis: document AI for invoices (m/f/d)

What it’s about

A crooked scan of a delivery note, an invoice with three pages of line items, a waybill with a handwritten remark: in many businesses, people still type documents like these into the ERP system field by field. Modern language and vision models can take that over – but how accurately, at what scan quality, and with how much checking afterwards? That is the question behind this thesis.

vona is an AI & software agency from Leipzig working remote-first for small and mid-sized companies, established businesses, trades and startups across the German-speaking countries. We combine AI integration, development, design and SEO in one team, and we give honest advice – even when the answer is “this isn’t worth it”.

For a logistics provider, we automated document processing: up to 400 delivery notes, waybills and customs documents a day run through OCR and a locally hosted language model. The result according to the client: 85% less manual data entry, 12 seconds instead of 4 minutes per document, and around 99% of the data extracted correctly. Your work can build on this. One possible research question: how do classic OCR pipelines followed by a language model compare with multimodal models that read the image directly – in terms of accuracy, cost and the kinds of errors they make? Your own ideas, for instance on table recognition or on models assessing their own confidence, are welcome.

Your tasks

• Put together an anonymized test corpus covering different document types and scan qualities, and define which fields need to be extracted.
• Build a pipeline, for example in Python with OCR, Ollama and a model such as Llama, and compare it with cloud models such as Claude, GPT or Gemini.
• Develop a method that lets the system flag uncertain fields and route them to a person for review.
• Evaluate accuracy per field, processing time and cost per document.
• Formulate recommendations on which approach suits which document volume and which data protection requirements.

What you bring

• Studies in computer science, business informatics, data science or a related field
• Basic Python skills and an appetite for experimenting with language and vision models
• Care in handling data and an eye for clean evaluations
• An interest in processes in logistics, retail or administration
• Good German, as the documents and our clients come from German-speaking countries

Nice to have, not a must: experience with OCR tools, PostgreSQL or workflow tools such as n8n.

What we offer

• Work close to real client projects, not made-up examples
• Subject-matter support from a dedicated contact at vona and regular check-ins; academic supervision is handled by your university
• Access to the AI tools you need for your experiments, such as Claude, ChatGPT, Gemini, Ollama and n8n
• Hours you plan around your semester, working hybrid – in Leipzig and from home
• If it goes well, a chance to keep working together, for example on follow-up projects – a possibility, not a firm promise

How to apply

Use the form at the bottom of this page to tell us a little about yourself and what interests you about document AI. A link to your CV, GitHub or LinkedIn is enough; there’s no need for a cover letter full of stock phrases. You’ll usually hear back from us within a week.