# First floor AI project card Fill in one card for the machine or defect detector you would fund first. Fill in what you know; every blank line gets an owner and a date. Any blank line means: not ready to fund. | Question | Answer | Owner / by when | | --- | --- | --- | | The expensive machine or fleet you would start with, and the failures you would predict | | | | Spares you would hold for those failures, and who holds them | | | | What an hour of its downtime costs, measured by whom and over what period | | | | What the model costs per year, all in: build, run, integration | | | | The test: does that downtime cost more than the model? (yes / no, with the two numbers) | ☐ Yes ☐ No | | | The other candidate: the defect or anomaly you would detect, and the historical data it trains on | | | | Its reach: what it can catch that you could not spell out rule by rule before | | | | If an analytics project stalled on this floor before, did a moved sensor or shifted use case force a restart? | | | | Baseline to prove the win: unplanned downtime hours a month, or defect rate, before anything starts | | | | Which candidate gets the first budget line, and why (one line) | | | | How it goes live with minimal disruption to production (window, line, discipline) | | | | Drift and performance monitoring: who watches it from day one, and what triggers a look | | | | Does the model touch a step that cannot be redone? (yes / no) | ☐ Yes ☐ No | | | If yes: the first-time-right validation it passed, named | | | | Quality bar on that step: highest missed-failure and false-alarm rate you accept | | | | Who can stop the model on the floor, and how fast | | | | Decision and date | ☐ FUND PREDICTIVE MAINTENANCE ☐ FUND DEFECT DETECTION ☐ NOT YET Date ______ Signed ______ | | ## Example (fictional): failure prediction on a bottling filler line Fictional example. | Question | Answer | Owner / by when | | --- | --- | --- | | The expensive machine or fleet you would start with, and the failures you would predict | Filler 2, the machine that fills and seals bottles; predict bearing and seamer-head failures | n/a | | Spares you would hold for those failures, and who holds them | Seamer heads and bearings held by the maintenance store; minimum stock agreed with Supplier B | Maintenance supervisor | | What an hour of its downtime costs, measured by whom and over what period | About $40,000 an hour of lost throughput; plant controller's estimate from the 2025 downtime logs | n/a | | What the model costs per year, all in: build, run, integration | About $180,000 a year for build and licenses; integration into the line controls not quoted yet | Engineering manager, quote by 30 Jan | | The test: does that downtime cost more than the model? (yes / no, with the two numbers) | **Blank: not ready**: integration cost missing, so the test cannot run yet | Engineering manager, by 30 Jan | | The other candidate: the defect or anomaly you would detect, and the historical data it trains on | Cap and label defects at the end-of-line vision system; two years of labeled images | Quality lead | | Its reach: what it can catch that you could not spell out rule by rule before | Cap skew and label wrinkles caught from the images, without writing a new rule per defect | Quality lead | | If an analytics project stalled on this floor before, did a moved sensor or shifted use case force a restart? | Yes: a 2023 downtime dashboard stalled after a sensor move forced a rebuild; restarts likely killed it | n/a | | Baseline to prove the win: unplanned downtime hours a month, or defect rate, before anything starts | 11 hours a month unplanned downtime on Filler 2 (2025 average); 1.4% defect rate at final inspection | n/a | | Which candidate gets the first budget line, and why (one line) | The filler prediction, because its downtime cost is the bigger number, once the test closes | n/a | | How it goes live with minimal disruption to production (window, line, discipline) | Shadow mode for 4 weeks with no line stoppages; go-live in the summer changeover week | Production planner | | Drift and performance monitoring: who watches it from day one, and what triggers a look | Plant data lead sends a drift and performance report to the maintenance supervisor every week, from day one | n/a | | Does the model touch a step that cannot be redone? (yes / no) | Yes: the vision system's auto-reject scraps bottles, and scrapped product cannot be redone | n/a | | If yes: the first-time-right validation it passed, named | **Blank: not ready** | Quality lead, validation plan by 12 Dec | | Quality bar on that step: highest missed-failure and false-alarm rate you accept | Missed cap defects above 0.5% on the audited sample unacceptable; false alarms above 5% trigger a review | Quality lead | | Who can stop the model on the floor, and how fast | Any shift supervisor can switch the line back to manual inspection; one switch on the line panel | n/a | | Decision and date | ☐ FUND PREDICTIVE MAINTENANCE ☐ FUND DEFECT DETECTION ☒ NOT YET · Date 20 Nov · "Test row open and no validation named for auto-reject." | Plant manager | From the course AI in manufacturing, Dr. Tony Hoang · v1 · Oct 2026