In 2018, a wildfire caused by a faulty electric transmission line owned by Pacific Gas & Electric (PG&E) tore through Northern California's Butte County, killing 85 people and destroying nearly 19,000 buildings. In June of this year, PG&E CEO and President Bill Johnson, on behalf of the utility, pled guilty to 84 counts of involuntary manslaughter and one felony count of unlawfully starting a fire in Butte County Superior Court.\nIn the wake of the fire, PG&E, which serves 5.2 million households in Northern California, set out to develop an AI technology suite that leverages computer vision to help it identify high fire-risk areas. Dubbed Sherlock Suite, the solution has helped PG&E automate inspections of its field equipment.\n\n[ Learn the essential skills and traits of elite data scientists and the secrets of highly successful data analytics teams. | Prove your data science chops by earning one of these data science certifications. | Get the insights by signing up for our newsletters. ]\n\n"The Sherlock Suite allows desktop inspectors to mark up potential equipment problems on high-resolution images, further training computer-vision models to automatically detect potential issues and adding metadata to enable searchability of these images across the enterprise," says Kunal Datta, product manager for the Sherlock Suite at PG&E.\nThe Sherlock Suite has earned PG&E a CIO 100 Award in IT Excellence.\nAutomating inspections\nAfter the fire, PG&E captured more than 2 million images of 50,000 electric transmission towers using aerial photography. It hired 150 desktop inspectors from around the country to review the images. The inspections were initially done using folders on a shared drive, with paper manuals, legacy map systems, and an Excel spreadsheet used to track the work.\n"The Wildfire Safety Inspection Program was the first time that PG&E conducted remote inspections using aerial imagery at this scale," Datta says.\nThe manual process suffered from long lead times from image capture to inspection, and the inspections themselves were time-consuming. In January 2019, PG&E formed the Sherlock team. The team interviewed inspectors, supervisors, subject-matter experts, leaders and various others across the inspection program to identify opportunities to streamline and automate processes.\n"Tracking the work from flight to inspection completion required manual entry of data throughout the process," Datta says. "Reducing wildfire risk is a priority for PG&E, and so reducing the time to inspection \u2014 as well as inspection time \u2014 and increasing auditability across the inspection process was identified as an important area for improvement."\nThroughout the development process, Datta says his team \u2014 consisting of data scientists, developers, data engineers, product management, and design \u2014 kept in constant contact with stakeholders to understand the issues from their perspective.\n"A key part of our philosophy is to collaborate closely with our business partners. We don't build stuff for our business partners \u2014 we build it with them," Datta says. "This level of engagement helps us come up with small, testable increments that we then get feedback on. The key to identifying the right thing to build, for us, is to make sure that we have a tight feedback loop with our users."\nIterate \u2014 and manage expectations\nThe team deployed a beta to a small group of inspectors in March 2019 and moved the entire inspector team to Sherlock by May 2019, though it continues to add features.\n"We're always building. There is no true 'done' state," Datta says. "We push new releases multiple times a week with small changes that we get feedback on constantly. We use Scrum, and so the team has a biweekly sprint review with all the stakeholders where we present what we did the last sprint, and what we're doing the following sprint, making sure to leave plenty of time for feedback and discussion afterwards."\nThe Sherlock web application enables inspectors to look at photographs and mark them with issues they find. The markups are used as labels to train computer-vision models, which then provide predictions to inspectors via Sherlock. The inspectors give a thumbs-up or thumbs-down to the prediction, further refining the models. The suite automatically flags standard items required for compliance review.\nBorrowing from the lingo established by the Society of Automotive Engineers (SAE) to discuss levels of automation in self-driving cars, Datta explains PG&E is currently in the process of transitioning from Level 0 automation (no automation, manual processes) to Level 1 automation (automated assistance). He notes that tempering expectations is a critical challenge.\n"When we say we are using artificial intelligence, people get excited. That's definitely a good thing, but it also means that expectations can be all over the map," Datta says. "When we think of AI, some folk jump straight to expecting Level 5 automation and ask when we will get there."\nDatta notes that he keeps a few slides called "Machine Learning 101" in every presentation he does to make sure that everyone understands that AI is math, not magic.\nThe Sherlock Suite has already dramatically reduced inspection time and time to inspection, and Datta says both metrics continue to improve as the team deploys new features. The suite has also allowed the Electric Operations organization to search for images, and Datta says other parts of the business are also gaining interest in the models as they start seeing new opportunities as a result of the capabilities engendered by Sherlock.