Startup idea: continual on‑prem learning agents

Industrial AI is getting a lot of attention in manufacturing right now, but one of the biggest questions is still the most practical one. How do you turn plant data, process knowledge, and operational constraints into something that actually creates value? In this episode of Manufacturing Hub, Vlad Romanov and Dave Griffith sit down with Konstantin Paradizov of Eukodyne for a detailed conversation on what industrial AI looks like when it is applied by people who understand manufacturing, MES, process improvement, data architecture, and the realities of the plant floor.

What makes this discussion especially valuable is that it does not stay at the surface level. Konstantin shares how his background moved from pharma into food and beverage, how Lean Six Sigma and process thinking shaped his approach, and why many of the best opportunities in manufacturing still begin with understanding the actual workflow before talking about software. The conversation explores a theme that comes up again and again in industrial transformation: the biggest gains often do not come from adding more technology first. They come from understanding the problem clearly, identifying what information matters, validating assumptions with the people doing the work, and then using the right mix of tools to move faster.

A major part of this episode focuses on the real use of AI in consulting and discovery. Konstantin explains how his team uses secure transcription workflows, on premises AI infrastructure, cloud models, masking of sensitive information, iterative validation, and ROI driven reporting to create high value outputs in a fraction of the time that would have been required even a year or two ago. This is an important point for manufacturers, system integrators, software teams, and plant leaders. AI is not just something that sits in front of an operator as a chatbot. It can be used behind the scenes to accelerate analysis, strengthen recommendations, shorten discovery, improve documentation, and reduce the cost of getting to a better answer.

The technical section of this episode is especially strong for anyone working in industrial automation, OT data systems, or applied AI. The discussion covers on premises compute, Nvidia based edge hardware, Linux environments, Docker containers, RAG workflows, vector databases, knowledge graphs, MQTT pipelines, HiveMQ, Mosquitto, n8n, Claude Code, Cursor, Gemini, OpenRouter, and the tradeoffs between frontier models in the cloud and smaller or open models deployed closer to the process. One of the clearest takeaways is that manufacturers should not start with the biggest model or the most exciting headline. They should start with the problem, the constraints, the data path, and the economics of the solution.

Vlad also pushes on an issue that matters to almost every manufacturer trying to prepare for AI. If you collect massive amounts of plant data into historians, cloud platforms, and enterprise systems, is that enough to create value later? Konstantin’s answer is thoughtful and realistic. More data alone does not automatically lead to better outcomes. You still need filtering, context, prioritization, architecture, and a disciplined way to separate signal from noise.

Learn more about Joltek here:

• https://www.joltek.com/services
• https://www.joltek.com/services/service-details-it-ot-architecture-integration

Connect with our guest:
Konstantin Paradizov
https://www.linkedin.com/in/konstantin-paradizov/

Learn more about Eukodyne:

• https://eukodyne.com/

Follow Manufacturing Hub for more conversations on industrial AI, digital transformation, OT architecture, SCADA, MES, industrial data strategy, systems integration, and the future of manufacturing technology.

Timestamps
00:00 Welcome and introduction to industrial AI applications
01:50 Konstantin’s background from pharma to manufacturing
05:30 Why food and beverage offered major process improvement opportunities
08:10 How to identify the right manufacturing opportunities to pursue
13:10 Using AI to accelerate discovery, documentation, and customer value
21:20 The on premises AI hardware stack and model selection strategy
30:10 Why iterative validation still matters more than a first AI answer
39:00 Claude Code, developer workflows, and practical AI tool stacks
48:20 On premises versus cloud AI and how to think about the tradeoff
54:10 Small models, low cost hardware, and edge deployment realities
01:05:00 Plant data, historians, filtering, and separating signal from noise
01:14:50 Predictions for industrial AI, career advice, and final recommendations

References and resources mentioned in the episode
MaintainX

• https://www.maintainx.com/

Solve for Happy
https://www.mogawdat.com/books

George Orwell 1984
https://www.penguinrandomhouse.com/books/326569/1984-by-george-orwell/

George Orwell Animal Farm
https://www.penguinrandomhouse.com/books/561805/animal-farm-by-george-orwell/