Paco Nathan argues that enterprise teams need help moving research and open-source projects into production, especially for natural language, graph, and machine learning applications.
2
Small, specialized models combined in ensembles can reduce cost and keep sensitive enterprise data inside the company's environment while supporting narrow, well-evaluated tasks.
3
Manufacturing AI often creates more value by extracting and checking information from PDFs, then supporting finance and value-chain decisions, than by starting with factory sensor data.
Summary
Paco Nathan describes his work helping enterprise teams turn machine learning research and open-source software into production systems. He explains why smaller specialized models and ensembles fit many enterprise tasks better than a single large language model. They can reduce costs, support privacy, and make evaluation more specific. Nathan also argues that data quality, domain expertise, and operations matter more than model benchmarks alone. He is especially concerned about experienced workers retiring from manufacturing, climate science, and other industries. AI can capture some of their knowledge through human-in-the-loop systems and active learning, although it cannot simply replace them. In manufacturing, he expected sensor data to dominate, yet he often finds the more valuable information in PDFs containing patents, scientific documents, shipping manifests, environmental reports, and compliance material. Extracting and validating that information can support finance and value-chain analysis. Nathan is candid that company-specific schemas, acquisitions, security, legal risk, and operational constraints make generic platforms a poor answer.
Enterprise adoption depends on turning research into usable production software
Paco Nathan says Derwen helps enterprise teams bridge the gap between newly published research and software that another team can install, reproduce, and use in production. Teams may find a promising paper and its open-source implementation, then still need to check whether it installs correctly, reproduces the reported results, and fits their use case. His work focuses on integration across natural language, graph technologies, and machine learning. Nathan connects this work to his earlier experience with data teams, AWS, Spark, graph algorithms, and open-source natural language projects such as pytextrank.
Hardware economics drive changes in software and team processes
Nathan argues that people often describe technology evolution backwards. Hardware changes alter what software can do, and that changes how teams work. He gives the example of a workload that once needed a large Hadoop cluster on EC2, then ran faster after being rewritten in Python on a laptop. As practices become common, teams tend to move away from heavyweight platforms and toward libraries used at the application layer. He expects large systems to remain for specific workloads, while many teams will get most of what they need from open-source libraries.
Enterprise AI is moving toward smaller models and model ensembles
Nathan points to talks by Waleed Kadous, Christopher Nguyen, and Mark Wang as evidence that enterprise teams are choosing narrow models and ensembles. Kadous showed that several specialized models can reduce the cost base by about an order of magnitude while preserving more privacy and security. Nguyen focused on selecting narrow tasks with good evaluations. Wang described combining signals from procurement, security, and legal teams before synthesizing an answer. Nathan says these approaches fit organizations that cannot send confidential data to a third party and want to inspect and run models within their own cloud perimeter or on premises.
Domain-specific evaluations matter more than model leaderboard results
Nathan says enterprise AI teams need their own data, evaluations, and domain experts. Public benchmarks often compare models while ignoring cost, security, and the details of a company's process. He describes this as a shift from a model-centric view to a data-centric view. Domain experts know what a correct result means for the business, so they need to help define the evaluation set and review the system. Without that work, Nathan says teams are placing applications on an unstable foundation, even if the selected model performs well on general benchmarks.
AI can preserve knowledge as experienced workers leave industry
Nathan says the loss of experienced workers is one of the concerns keeping enterprise AI leaders awake. In areas such as sustainable fishing, steelmaking, climate science, and chemical manufacturing, people may need many years to learn how to recognize problems and choose the right response. He gives the example of Furuno sending people onto Japanese fishing boats with laptops to build localized models. Under Japanese law, the fishermen own the models. The goal is to capture their knowledge and give younger workers a form of mentorship, rather than claiming that AI can simply replace the workforce.
Human oversight is necessary for dangerous industrial decisions
Nathan describes a steel-mill worker who stands near molten steel and listens for signs that the furnace needs maintenance. A mistake in either direction can create a large financial loss, while training someone to make that judgment can take about ten years. AI could monitor sounds and route information to an expert, reducing the need for every expert to stand beside a furnace. Nathan still insists on human involvement, active learning, and weak supervision. The system should bring experienced workers into the process instead of treating the task as ordinary automation.
AI systems need software engineering, operations, security, and legal review
Nathan rejects the idea that an AI engineer can succeed by calling a model API without broader systems knowledge. He says operations is often the bottleneck in manufacturing because companies struggle to hire and retain strong operations talent. AI applications also need software engineering for concurrency, caching, data processing, and workflow orchestration. Teams must assess outages, model changes, liability, security, and regulatory compliance. Nathan describes legal professionals and software engineers working together at a law firm to examine these risks. A model call is only one part of an application with many dependencies.
Manufacturing value often sits in PDFs and business data
Nathan expected manufacturing use cases to center on sensors, streaming data, and equipment monitoring. In practice, he often finds more immediate value in finance and value-chain analysis. Important information can be buried in patent applications, scientific documents, environmental reports, shipping manifests, and regulatory material, much of it stored in PDFs. PDF parsing is difficult because text and graphics are arranged as separate boxes that must be reconstructed. Extracting and validating this information can help companies understand vendors, competitors, compliance obligations, and the investments behind a customer order.
Generic platforms struggle with company-specific schemas and acquisitions
Nathan says a platform cannot solve every enterprise problem because processes differ across companies and business units. Large organizations may have acquired many teams, each with its own database and working practices. Understanding and aligning those schemas can become a major operational problem. He mentions an aircraft manufacturer that took database systems offline for almost a month each year to perform schema alignment after many acquisitions. For Nathan, the practical approach is to work with individual business units and their domain expertise instead of promising a universal platform.
"Trying to come in from the outside and say hey I've got a platform is going to solve all your problems, it's laughable because there are so many problems and they're so different and they change from company to company."Paco Nathan1:15:59
Who should watch
You are building an enterprise AI application and need to decide whether a large hosted model is justified.
Your team is working with manufacturing, finance, compliance, or other domains where valuable knowledge sits in documents and with experienced workers.
You need to plan evaluations, operations, security, and legal review before putting an AI system into production.
