MLOps Capstone: Design and Defend a Production System

An electric utility wants to prioritize inspections of poles and vegetation using aerial images, weather, maintenance history, and sensor data. The ranked list will influence field crews, emergency preparation, and maintenance budgets.

The inspection regime that list feeds into is not an informal habit, and it is not yours to define. In California the rules it can be judged against carry numbers. On 26 November 2019 the Safety and Enforcement Division of the California Public Utilities Commission moved to widen its investigation to cover the 2018 Camp Fire, and attached a table of the violations it had found: General Order 95 Rules 18, 31.1, 31.2 and 44.3, General Order 165 Section IV, Resolution E-4184, and Public Utilities Code § 451. One of them was the failure to conduct detailed climbing inspections when internal triggers were evident.

What an inspection coverage gap looks like in that regime is also on the record. The Camp Fire ignition point was Tower 27/222 of PG&E's Caribou-Palermo 115kV line, whose transposition hardware had been in service since 1921. The Butte County District Attorney, Michael L. Ramsey, worked through every inspection and patrol record he could obtain back to 2001. His 92-page public report of 16 June 2020 found nothing there at all: “There is no record of any climbing inspections, detailed ground inspections above 10’ or aerial inspections conducted on the Caribou-Big Bend section of the transmission line.” The same report documents that PG&E had “little or no information about the 97-year-old conductor”, and that a 1995 ES Guideline had eliminated routine climbing inspections.

That is the ground your ranked list lands on: a numbered standard, a documented coverage failure, and a regulator willing to enumerate which rule each gap broke. Labels are delayed and incomplete. Remote regions have weak connectivity. Wildfire season creates sharp demand peaks, and the company must explain its decisions after incidents. So the capstone is to design the operating system around the model, not to choose the most impressive architecture.

The capstone should produce artifacts, not only an architecture diagram. Each packet below exists because some real review — a district attorney's investigation, a regulator's violation table, a court's footnote — has already asked for exactly that document and found it missing somewhere.

The proposal should connect each layer to a requirement and recovery path: evidence and learning, release control, execution, operations, and governance and lifecycle.

The governance layer is the one teams most often draw as an empty box. It is also the one with a statute already written into it. The AI Act, Regulation (EU) 2024/1689, was published in the Official Journal on 12 July 2024 and entered into force on 1 August 2024. The European Parliamentary Research Service gives the citation as “Regulation (EU) 2024/1689, OJ L, 2024/1689, 12.7.2024”. Its Annex III, the list Article 6(2) points to, names at point 2: “Critical infrastructure: AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity.”

A model that ranks inspections for an electricity supplier is therefore not in a governance vacuum, and “governance” is not a box to be filled in later. The classification carries obligations on risk management, data governance, human oversight and post-market monitoring, before and after the system is placed on the market. Each of those obligations should terminate in a named artifact somewhere in your architecture graph, with an owner. Not in a colored rectangle.

The project is not required to end with full automation. Three positions can be defended: decision support only, constrained automation within a narrow eligible scope, and no deployment yet.

The gap between the first two and the third is measurable, and one agency measured it on its own program. On 11 August 2017 Michigan's unemployment agency published a review of 62,784 unappealed fraud-penalty cases from October 2013 to August 2015. Where the MiDAS computer program had resolved the case by itself, the finding was reversed 85 percent of the time: “Of those cases, 40,195 were originally resolved by way of computer program based on available information. As part of the review, 85 percent of these original fraud findings were reversed.” The other 22,589 cases were initiated by the computer and then reviewed by an investigator. Those were reversed at 44 percent. Same program, same period, same population — 85 percent against 44 percent, and refunds of more than $20.8 million.

The Michigan Supreme Court put the human cost of the automated branch on the record in Bauserman v Unemployment Insurance Agency, decided 26 July 2022. Roughly 40,000 people were wrongly accused, and footnote 5 notes that “a study conducted by the Agency concluded that, during this same period, approximately 93% of the automated system's fraud determinations were incorrect”. A reviewer in the loop is not a rhetorical concession in a design memo. In the only place anyone measured it here, it halved the error the system shipped.

Confirmed failure and fire outcomes are rare, while inspection findings are selected by previous policy. Weather, sensors, contractors, and image sources change. The system therefore needs proxy monitoring, independent audits, mature outcome cohorts, and explicit limits on what the model can claim.

What an independent audit does to a deployed model's own numbers is documented. The Epic Sepsis Model was in use at hundreds of US hospitals when Wong and colleagues validated it from outside, in JAMA Internal Medicine in 2021. Across 38,455 hospitalizations of 27,697 patients at Michigan Medicine, the hospitalization-level AUC was 0.63 (95% CI, 0.62-0.64). The vendor's own documentation said 0.76-0.83. The operational shape of that gap is in the abstract: “The ESM also did not identify 1709 patients with sepsis (67%) despite generating alerts for an ESM score of 6 or higher for 6971 of all 38 455 hospitalized patients (18%), thus creating a large burden of alert fatigue.” Their conclusion was that “the ESM has poor discrimination and calibration in predicting the onset of sepsis”.

One audit can be argued with. A second one, at different institutions, is harder. Ostermayer and colleagues ran the same model against two county emergency departments and reported in JAMIA Open in 2024 that “Within a 6-hour time window for sepsis, the ESPMv1 had a sensitivity of 14.7%, specificity of 95.3%, positive predictive value of 7.6%, and negative predictive value of 97.7%.” Two independent groups, one already-deployed model, and figures the vendor's documentation did not lead anyone to expect.

A valid conclusion may therefore be to deploy only as decision support, restrict certain regions, or retain a rules-based baseline until evidence improves. The evidence gap is not resolved by shipping and watching a dashboard. It is resolved by someone outside the build measuring the thing.

Each pass produces an artifact that can be reviewed independently: the product and risk contract, the asset and architecture graph, the evidence and experiment plan, release and recovery, operations and incident response, security and governance, and the decision memo.

Pass 4 is the one teams treat as plumbing, and it has the sharpest documented failure. Knight Capital Americas rolled new RLP code out to eight servers. One server did not get it and kept running discontinued Power Peg code. The SEC's order of 16 October 2013 sets out how: “During the deployment of the new code, however, one of Knight’s technicians did not copy the new code to one of the eight SMARS computer servers. Knight did not have a second technician review this deployment and no one at Knight realized that the Power Peg code had not been removed from the eighth server, nor the new RLP code added. Knight had no written procedures that required such a review.”

On 1 August 2012 that eighth server turned 212 incoming parent orders into 4 million executions in 154 stocks, for more than 397 million shares, in about 45 minutes. The loss was $460 million. Knight paid a $12,000,000 civil money penalty. Carol Clark, writing for the Federal Reserve Bank of Chicago, named the control that was absent: “Trading firms should have access controls defining who can develop, test, modify, and place an algorithm into production, as well as quality assurance procedures for the code and development processes.” A canary and a rollback plan in pass 4 are not ceremony. Here the entire failure was one machine out of eight that nobody was required to check.

The decision memo should compare alternatives, identify the safest useful scope, and state why the chosen complexity is justified. It should carry expected benefit, residual risk, evidence gaps, owners, stop conditions, and retirement criteria.

Approval is not the end of the obligation, and in this domain that is now written into law rather than into good practice. For high-risk systems under Regulation (EU) 2024/1689, the European Parliamentary Research Service records that once a system is placed on the market “providers must implement post-market monitoring and take corrective actions if necessary”. A memo that ends at launch has answered a question nobody with authority asked.

A recommendation not to deploy can earn full credit when it follows from the evidence. Michigan's own review of MiDAS, the two external validations of the Epic Sepsis Model, and the Nimrod Safety Case with 40% of its hazards still open are three different institutions discovering, after the fact, what a pre-deployment memo could have written down in advance. MLOps maturity includes the ability to prevent an unjustified system from becoming operational.