Overview
I spent July to September 2026 as a controls engineering intern at Jubilant HollisterStier (JHS), an FDA-regulated pharmaceutical manufacturer. I prototyped Line 2 SCADA screens and built a Microsoft Copilot Studio agent over more than 20 One Point Lessons — and both came down to the same issue: the information people needed already existed, but the path to it was long enough that nobody took it.
Problem
The SCADA screens on the line held everything an operator needed, but routine metric lookups took nine clicks to reach. A number that takes nine clicks is a number nobody checks mid-shift. Many of the screens were also visually dull and, in places, carried almost no usable information — a lot of surface area showing very little. Separately, maintenance information was difficult to reach in an existing job-aid library.
Constraints
- Regulated manufacturing environment; the SCADA screens were delivered as prototypes
- Operators won't be trained on my design; it reads at a glance or it fails
- Work within an existing industrial controls environment
- Eight weeks, including learning the HMI platform from zero
- An agent that confidently invents a procedure in a plant like this is worse than no agent
Architecture
I built the HMI prototypes in AVEVA InTouch and integrated process historian data from Allen-Bradley and Siemens PLCs to validate behavior against real production equipment. I also deployed software on FacilityPro and supported controls engineers on legacy manufacturing lines. The Copilot Studio agent retrieved and summarized more than 20 One Point Lessons, with explicit no-answer fallbacks and interoperability with a manufacturing intern’s agent.
Decisions
Dashboards split by role
A line worker and a manager watch the same process for different reasons. One needs to know what their equipment is doing right now; the other needs to know how the line is tracking. A single shared screen gives both of them a compromise. Separate dashboards let each view be built around the decision the person in front of it is actually making.
Real historian data in the prototypes
Mocked values prove a layout renders. Historian data proves the screen works against how the plant actually behaves — and the argument makes itself when the numbers on screen are the ones the reviewer already recognizes.
Explicit behavior when the answer doesn't exist
The agent's most important case is the one where the library has nothing. When a job aid describes the correct way to work on production equipment, a plausible invented answer is a safety problem rather than an inconvenience. Defining the gap cases was most of the real work.
Handing off between agents instead of dead-ending
I worked with an intern on the manufacturing side to get our two agents talking to each other, so a question that landed in the wrong place still got somewhere useful instead of returning nothing.
Challenges
Usability and efficiency pull against each other
The brief was to make the screens easier to read without losing efficiency, and those usually trade off — making something more legible generally means showing less of it. Most of the work was deciding what earned a place on the first screen and what could be one level down.
Learning an industrial HMI platform in eight weeks
I started with no AVEVA InTouch experience and had to be productive in it fast enough to ship prototypes worth reviewing, while also supporting the controls engineers on legacy lines and equipment issues around the plant.
Results
- Routine metric lookups went from nine clicks to three — I timed the old path myself, tested the new one, then retested
- The prototype screens led with KPI numbers, production timelines, and alarms
- I built separate prototype dashboards for different roles
- The Copilot Studio agent retrieved and summarized more than 20 One Point Lessons
- Document retrieval went from an average of seven clicks to a single query
- I implemented no-answer fallbacks and cross-agent interoperability
What I'd do differently
- I split the dashboards by role based on how people described their work, not on watching them do it. The screens stayed prototypes, so I never found out whether I drew the line in the right place — whether a line lead needs the manager view, for instance. I'd sit on the floor through a shift before deciding who sees what.
- I would track sustained use of the maintenance agent alongside document retrieval effort.