Slater Victoroff argues that data, rather than model architecture, is now the main limit on production machine learning.
2
Multimodal models are needed for documents because layout, position, tables, rich text, and images carry information that OCR text alone loses.
3
Machine learning systems should put subject matter experts close to supervision while data scientists design experiments and maintain the information flow.
Summary
Slater Victoroff describes how his early work with traditional machine learning led him to focus on making deep learning usable in enterprises. He defines unstructured data as text, audio, images, and combinations of them, including documents joined with metadata. For documents, text alone is insufficient because layout and visual structure affect meaning. Indico Data uses multimodal techniques to combine language with positional and visual information. Victoroff argues that modern models are capable enough that data quality, consistency, and process design have become the harder problems. He explains why changing an OCR engine can invalidate labels, and why labels should preserve both text and document position. He also discusses machine teaching, synthetic data, active learning, and human overfitting. His practical advice is to define the problem well, create consistently labeled data, involve subject matter experts, and treat production ML as a large enterprise system rather than a solo modelling exercise.
Deep learning pushed Victoroff from model competition toward enterprise access
Victoroff recounts that he began with traditional methods such as TF-IDF, logistic regression, support vector machines, and keyword features. While working on increasingly ambitious Kaggle projects with Alec Radford, he saw modern convolutional neural networks pull ahead in a tangible way. After trying to beat deep learning with further research, he accepted that he could not. He then chose to focus on making the technology accessible and enterprise-grade. That decision led to Indico Data.
Unstructured data includes every non-tabular modality and their combinations
Victoroff defines unstructured data as everything that is not rows and columns. That includes text, audio, images, video, documents, and combinations of these sources. Structured metadata attached to an unstructured source, such as an author field on a PDF, does not remove it from the unstructured use case. He also explains that the term contains nested distinctions. A fixed-position form may be structured within the wider unstructured category, while a long contract depends much more heavily on language.
Adoption depends on making advanced methods usable beyond narrow vertical systems
Victoroff says adoption has been difficult because simple techniques work well for tightly constrained documents, while modern language methods require a different infrastructure and modelling approach. Many successful applications have therefore appeared inside specialised companies that spent years building one narrow model, such as legal analysis. As the technology improves, companies are beginning to see unstructured work as a broader problem that can be handled with a unified stack rather than many isolated systems.
Documents need visual and positional information alongside language
Victoroff says a document is not simply text. People who process invoices or read contracts can lose the ability to do their work when visual formatting is removed and only OCR text remains. Tables are a clear example because nesting and reading order cannot be recovered through language modelling alone. Multimodal systems combine a language-model backend with local visual information, document position, bundle position, and other metadata. Indico uses a technique he calls model surgery to add this information.
Unstructured ETL breaks when labels are tied only to a particular text rendering
Victoroff gives an OCR and extraction workflow as an example. A document may be OCRed, classified, and passed to an extraction model that creates a structured payload. If labels are stored only as positions in the OCR string, changing the OCR engine can change the text and reading order. The labels then cannot be reused, so the team must retrain and relabel. Victoroff says labels should preserve both the document string and the location of the information, allowing them to be converted into the new text representation.
Machine teaching puts human reasoning closer to the supervision process
Victoroff introduces machine teaching as a way to design supervision around how people actually reason. He uses binary Twitter sentiment as an example of a poor framing. More agreement will not solve the limits of forcing sentiment into positive or negative categories when the real issue is how the task is defined. A richer form of supervision could make reasoning explicit, such as marking positive and negative aspects in aspect-oriented sentiment. The subject matter expert often has the information needed to supervise the model.
Synthetic data cannot create information that the original process did not contain
Victoroff compares synthetic data to Maxwell's demon, whose apparent creation of free energy fails because the information processing has a cost. In the same way, generating more examples from a model does not automatically add knowledge. Feeding synthetic examples back into training can reinforce what the model already believes. He makes a related argument about active learning. Generic uncertainty sampling can sharpen decision boundaries, but it cannot tell whether the boundary itself is wrong. Human choices about which examples are genuinely unusual add information to the process.
Production systems need staged updates and clean information flows
Victoroff says the person who understands the process best is often best placed to train the model, while the data scientist should design experiments, choose useful metrics, and maintain the system. Indico uses a staged workflow in which corrections feed a staging model rather than immediately changing production. Teams can review the staging model and opt into an update. This avoids an uncontrolled closed loop, which Victoroff calls a compliance problem and a path to damaging the model without a recovery method.
Data architecture and problem definition matter more than endless model tuning
Victoroff advises engineers to start with a well-defined problem and good, consistently labeled data. He says teams often want machine learning to act like an oracle, but they would have no way to know whether it was correct without a sound evaluation process. He also warns that a demo can be easy while taking years to reach production. Enterprise systems require coordinated help because they involve difficult data, compute, process, and compliance decisions. His view of the future data scientist is someone who designs data-intensive systems with strong statistical judgement.
