Co-founder project · Construction
Estimate AI
A model per vendor, fine-tuned from examples, then used on new PDFs.
- AI
- Web
Proof: Demo
Co-founder project · Document AI
Project Parse
Invoice extraction that retrains itself on every correction.
System diagram
- 01
Invoices
PDFs in any layout
- 02
Extraction
Fine-tuned Gemini 2.0 Flash
- 03
Review
Corrections captured
- 04
Retraining
A new model from corrected output
- 05
JSON
Fields to your schema
- Input or existing system
- We built
- Human step
- Built by
- Muhammad Zubair, Sortup co-founder
- Sector
- Document AI
- What was built
- AI
- Stack
- Gemini 2.0 Flash
- Gemini fine-tuning
- Python
- Django
- Celery
- Redis
- GCP
- Client
- A finance operations team processing supplier invoices.
- Engagement
- Zubair built the extraction pipeline and the fine-tuning loop around it.
Proof you can open
The challenge
Invoices arrive in every layout, and retyping vendor, line items, dates, totals and tax by hand is slow and error-prone.
The solution
Processed documents are converted to images and uploaded to GCP, where a fine-tuning job on Gemini 2.0 Flash trains a new model. When a model gets a document wrong, the corrected output goes into the next round. Django, Celery and Redis run the jobs.
The impact
- Fields extracted to a custom JSON schema.
- Corrections feed the next fine-tuned model automatically.
- Runs as scheduled jobs, not a manual process.
Building something like Project Parse?
No pitch deck. An engineer, and a straight answer.