Data-Driven Manufacturing:

AI, ML, and Data Science

Dr. James A. Bednar
Director, Custom Services
Manufacturers are awash in data!
• Sensors are cheap – you probably have
loads of them already!

• Plus you have lots of data on your

inventory, supply chains, and markets

• Making intelligent decisions based on your

data can give you a major leg up

• McKinsey estimates that $1-2 trillion in

manufacturing revenue is on the table

© 2023 Anaconda 2
“Manufacturing” covers a huge range of activities

• High-volume, high-tolerance mass production of commodities:

Need data to reduce material and energy costs, increase yields, predict maintenance

• Low-volume, low-tolerance precision manufacturing:

Need data to detect defects, improve performance, optimize processes

• Mass customization:
Need data to drive computer-controlled machinery and processes

• Fully custom shop:

Need data to meet specs and determine price points, feasibility

No single approach is going to meet all these needs!

© 2023 Anaconda 3
So… how do you extract value from your data?
If you’re Facebook/Amazon/Apple/Netflix/Google/Microsoft, you

• hire world-leading data-science and machine-learning/artificial-intelligence experts

• push the boundary of what is possible for computers and software to achieve
• build on nation-state-level data infrastructure

Is that the right approach for you, though?

© 2023 Anaconda 4
Don’t hire a big, expensive AI/ML/DS team with grand goals
• UNLESS you already have extensive data infrastructure in place
• AND have an extensive team of subject-matter experts (SMEs)
• AND the SMEs are motivated and officially charged with making this project a success
• AND you have the power to rebuild the organization’s decision-making processes to
be driven by data
Without all those in place, success at ambitious goals is highly unlikely.

© 2023 Anaconda 5
Don’t just throw ML at your problem!
• You can find all sorts of online tutorials and “auto ML” tools that promise instant
insights from arbitrary datasets

• Across many customers, industries, and projects over many years, we are not aware of
that promise ever having been realized, even with newer LLM/generative-AI models.

• Main problem: not understanding the data, the underlying system, and how they relate
• Insight from a raw dump of numbers is very unlikely; what’s signal and what’s noise?
• Conversely, domain experts don’t typically have the data-handling, interpretation, and
coding skills needed to work with large and complex datasets

© 2023 Anaconda 6
Instead, make a small, feasible plan
• What is your goal? Be very specific about your desired outcome.
(Hint: Pick something feasible to start with, e.g. improving one particular process)
• Is this a one-time need, or ongoing?
(Hint: Start with a one-time need until you get up to speed; data pipelines are hard)
• Do you already have the data you need?
(Hint: Start over with a different problem, if you don’t yet have data for the one you want)

• Do you have the organizational will to make changes?

(Hint: Use a small project to demonstrate results and get buy-in, first!)

• Do you have people who understand the process/system AND how to work with data?
(Hint: It won’t work if you don’t, but they don’t have to be the same people.)

© 2023 Anaconda 7
How Anaconda helps manufacturers
• Anaconda as a company is built around open source (OSS) Python libraries:
– Packaging OSS tools
– Helping you building OSS-based environments for computations
– Expanding and advancing OSS tools for working with data

• We help companies establish robust OSS-based computational pipelines

from raw data to visualization/analytics tools to automated processes
• The first step in any of our projects (and often the only) step:
Let’s take a look at (all) that data!

© 2023 Anaconda 8
Example Anaconda/manufacturing projects (1/4)
Consumer beverage company analyzing production samples

Issue: “poor performance” from ML model of chemical makeup of sample from

measurements

Solution:
• Visualized a set of black-box algorithms with HoloViews
• Found data being corrupted in compiled code early in the pipeline
• Simple non-ML algorithm sufficed after fixing that

© 2023 Anaconda 9
Example Anaconda/manufacturing projects (2/4)
Oil and gas company making production decisions for wells from sensor data

Issue: data pipeline too slow to handle data volumes

Solution:
• Numba to compile Python to fast machine code
• Implemented efficient chunked data handling for incoming streams of measurements
• Panel app displaying data for decision makers

© 2023 Anaconda 10
Example Anaconda/manufacturing projects (3/4)
Consumer device manufacturer optimizing for device performance

Issue: too much benchmarking data for standard visualization tools

Solution:
• Server-side rendering using Datashader to handle large data volumes
• Extensions to Bokeh and HoloViews to display many sources of information together
• Resulting analyses packaged as Panel dashboards

© 2023 Anaconda 11
Example Anaconda/manufacturing projects (4/4)
Construction materials manufacturer looking to improve processes

Issue: Disconnect between SMEs with problems to solve and data scientists with skills

Solution:
• Set up OSS tools for working with data across the spectrum from:
– non-coder SMEs (deployed dashboards and reports) to
– analysts (low-code dashboards and preconfigured notebooks) to
– programmers/data-scientists (Python)
• All tools connect directly to the same live data feeds
• All tools can share the same viz/analysis components whether full/low/no code
• Tools collect annotations from end users and display for humans or record for ML

© 2023 Anaconda 12
Tools we use in our client work
OSS:

• HoloViz.org: Visualize all your data (always the first step!)

• Panel.HoloViz.org: Use Python to build data apps for use throughout your organization
• Github.com/Panstacks/Pandata: Scalable data-processing pipelines
• MLFlow: Keep track of all your experiments and model runs
Commercial:

• Anaconda Distribution: Secure builds of all the Python OSS packages you need
• Anaconda Data Science Platform: Kubernetes-based AI/ML/DS development and
deployment, for building fully OSS data-processing pipelines (no vendor lock-in!)

© 2023 Anaconda 13
The Pandata Stack: Scalable Open-Source Analysis
Data Data Data Data User
Visualization
storage access API processing interface

Your domain-
specific code

Legacy/
domain-
specific
formats

Packaging

© 2023 Anaconda 14
Ok, I started small and it worked! Now what?
• Great! Do you have the organizational will to expand?
(If not, stay small while you work on the politics, consolidating your power)
• Do you have the data you need to expand?
(If not, is getting usable data feasible? It will always be more expensive and take longer
than you think!)
• Are there lots of other opportunities on specific processes?
(If so, see if you can set up tooling that can be used across the board)
• Or are you ready to start optimizing the whole factory’s operations?
(You’ll need even more buy-in across the organization!)
• Or are you ready to look outwards towards the whole supply chain?
(Understanding inefficiencies, gaps, and comparative advantage could be a big win!)

© 2023 Anaconda 15
Thank you and Q&A
• Thanks to all our amazing manufacturing
clients over the years!

• Special thanks to our partners at the

National Manufacturing Institute of
Scotland, especially Prof. Andrew
Sherlock, for helpful discussions!

• See the Anaconda Blog for more info

• Contact [email protected] or
[email protected] if you want our help
applying OSS to your own data problems!

© 2023 Anaconda 16