GNOSIS
Exploring OCR and Visual Language Models in the context of complex graph comprehension. Building a modular architecture that can swap out models as well as evaluate their performance on graphs of oil and gas wells.
Problem & Impact
Solves the problem of Multimodal and OCR models not comprehending graph and tabular content. OCR models are good at reading text, even when its blurry or grainy, but reading grainy and at times even hand-drawn graphs is even more tricky.
Google's Gemini has become the defacto model for graph comprehension, but as many have noticed, it easily gets pricey. With this said, we believe a potential combination of OCR-models, Visual Language Models and sophisticated scaling, resizing and skewing would be even more efficient.
Tech Stack
Key Features
- Run inference on graphs and plots via API, pick from any of the best benchmarking open source models.
- Get insights into the model performance on your particular media, to see which one would be optimal.
- Image Preprocessing, which includes denoising and unskewing.
Screenshots
Dashboard showing the modular design for model evaluation
GNOSIS architecture overview showing the modular design for model evaluation
Timeline
Start Date
November 2025
Current Phase
In development
Upcoming Milestones
- Deployment
- Evaluation suite
- Public API Access
Affiliations
Wellvector
Maintenance & Deployment
The goal is to deploy as pay-per-use with no profit, publish as docker image and separate evaluation components into an eval suite for visual AI.
Contributors
Niklavs Visockis
Tech Lead
Georg Zsolnai
Backend
Michael Yu
Backend
Sebastian Schmülling
Inference
Elias Lindstenz
Inference
Giulio Altomari
Fullstack