Zurück

ELIO: A Modular, AI-Driven Pipeline for Automated Label Digitization

Extracting structured data from chaotic physical media is a major bottleneck for digitizing biodiversity collections. The ELIO project (Museum für Naturkunde Berlin & Birds on Mars) solves this by combining Computer Vision, OCR/HTR, and LLMs to parse raw specimen label images.

We detail handling GPU limits via staged execution, using dynamic routing with resilient fallbacks, and building an automated QA engine that validates outputs against external databases to mitigate hallucination risks. Discover a practical blueprint for embedding complete data provenance directly into production outputs.

Vorkenntnisse

Basic knowledge of ML (LLMs and Computer Vision).
Familiarity with data engineering (APIs, pipelines, JSON).
Awareness of AI deployment challenges (GPU limits, hallucinations).
No deep MLOps or biology/museum background required.

Lernziele

Architect multi-modal pipelines combining CV, OCR, and LLMs for unstructured data.
Prevent GPU out-of-memory crashes via staged execution.
Build resilient MLOps using dynamic routing and fallback loops.
Automate QA by validating outputs against external databases and embedding strict data provenance into JSON exports.

Speaker

Margot Belot is a digital humanities and museum data specialist at the Museum für Naturkunde Berlin. She develops AI-assisted workflows for digitizing natural history collections, focusing on reproducible, open, and critically reflective practices. Her work spans FAIR data management, OCR/HTR pipelines, citizen science, and methodological transparency in cultural heritage digitization.
LinkedIn

Nicolas Gorges (Head of AI, Birds on Mars) leverages over 20 years of R&D experience in ML, CV, and robotics. He focuses on designing scalable AI architectures and transitioning state-of-the-art research into production systems. His tech portfolio ranges from experimental NLP/audio synthesis and predictive modeling to automated CV pipelines for digital archiving.
LinkedIn