What is AMI? Specifically, it stands for Advanced Metering Infrastructure, but over the past year or two, it has started to take on a bigger meaning. AMI is no longer just about collecting meter reads but instead is about connected solutions and devices transmitting large amounts of data over the network. The Sensus FlexNet system allows utilities to not only deploy smart meters like the Sonix IQ, but also allows utilities the ability to deploy devices like the Sensus Smart Gateway and Cathodic Protection Test Points that give utilities greater insight into pipeline integrity.

With the deployment of Sonix IQ meters, utilities will have more granular consumption data while also gaining access to pressure and temperature at the meter. The Sonix IQ will also have remote shutoff capability as well as automatic shutoff ability when pressure, flow, or temperature crosses a predetermined threshold. Outside of just consumption data, the Sensus Smart Gateway enables utilities to send pressure, temperature, position, or any piece of data that comes from a 4–20mA sensor over the FlexNet network. It can be deployed anywhere in a utility’s network and moved around as needed. The cathodic protection test points collect protective voltage data at cathodic protection test stations and sends the data over FlexNet for utilities to analyze and respond when voltage crosses a set threshold.

Utilities also have access to more data than ever before coming from devices, sensors, internal databases, and external datasets. Urbint supercharges utilities’ operational, inspection, and GIS data with environmental and external metadata from the locations where each facility is installed. That data is used to train Artificial Intelligence models using non-linear machine learning techniques, identifying patterns and relationships between features that are used to determine risk at each facility.

The Urbint Lens for Leak Prediction uses results of the AI predictions to show where leaks are most likely to occur, and to determine the expected number of identified leaks, by grade, across the broader service territory. The Urbint Lens for Main Health uses the same techniques to isolate the main segments most at risk from different threats. The utility can then use the software to understand the tradeoffs across safety, reliability, and affordability in their main replacement programs.

Before AMI, IoT, or whichever buzzword you choose to use, more data generally meant higher spending in SCADA, more trips to the field, and making sure that you had AC to power your devices. Data sources were typically disparate as no database or technology was connecting them, leading to many missed opportunities for improvements in operations. With recent advances in technology, utilities now have access to more frequent data at a fraction of the cost and can connect devices in ways previously unattainable.

With access to data becoming easier every day, it can be overwhelming and take years for a utility to integrate it into their analytics and data strategy. That’s where Sensus and Urbint come in. Marrying the Sensus and Urbint technologies enables utilities to connect new types of data to unlock novel insights. Complex interrelationships across large combinations of features can be calculated quickly using AI, for example allowing a utility to use more frequent consumption data to help predict methane loss on a leak by leak basis. Collecting higher resolution real-time data adds the ability to understand risk as it happens, and each new device that gets connected increases the predictive power.

For more information on Urbint check out our website and blog, and learn more about Sensus at their website and blog!

Urbint and Sensus — taking AMI beyond meter reading was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

Last week Urbint ventured south, to the Lone Star State, to participate in an exhilarating exchange of ideas and information with natural gas industry experts at SGA’s Natural Gas Connect in San Antonio. In true Texas fashion, there were big armadillos, line dancing and even bigger opportunities!

During the conference, Urbint showcased its AI powered Damage Prevention, Cross Bores, and Odor Call Forecasting solutions, which empower utilities to quantify trade-offs between safety, reliability, and affordability while reducing risk.

Benjamin Berry, Head of Product at Urbint, co-presented, “Data Analytics and Reporting: Transforming Data into Actionable Intelligence,” with Nicor Gas. Nicor Gas, one of the largest natural gas utilities in the Nation with 2.2 million customers, partnered with Urbint to take a risk and data-driven approach to address the three root causes of damage: excavator-error, locate-error, and no one call tickets. Together we demonstrated the impact of Urbint Lens for Damage Prevention to reduce third party damages for Nicor Gas.

The event also featured roundtable sessions on asset management, public safety, field operations, and emergency response with gas distribution operation managers from across the country.

A notable session on best practices to decrease leak response times was very insightful and included operation managers sharing successes and challenges to effectively respond to customer odor calls. As deployment strategists at Urbint, we focus on how making tradeoffs plays a significant role in managing response times and customer experience.

The conference unequivocally was a great chance to learn and exchange with knowledgeable industry professionals in one place. If you didn’t make it to this year’s Natural Gas Connect and would like to learn more about our AI-powered solutions or Urbint partnership opportunities, contact us!

Varun Adibhatla is a Deployment Strategist at Urbint. Varun holds Master’s Degrees in Information Science, from Penn State University, and Applied Urban Science and Informatics from New York University’s Center for Urban Science and Progress.

Out on the Range: A Recap of SGA’s Natural Gas Connect 2019 was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

Come chat with me (left) about how artificial intelligence (AI) can help gas operators maximize risk reduction at the Urbint booth (#807) during Southern Gas Association’s Natural Gas Connect Conference in San Antonio from July 15, 2019 to July 17, 2019.

In the last few months, Urbint has analyzed how states perform against damage prevention metrics, discussed how AI can be used to predict disasters to improve resiliency, and helped gas operators manage their operational risk.

I am looking forward to continuing these conversations with folks in person at SGA’s Natural Gas Connect, a conference to connect people, ideas, and information in the natural gas industry.

I’m especially excited to see SGA increase the number of roundtable sessions by 68% — from 19 the previous year to 32. Topics like “Dispatch & Field Resource Planning” and “Resource Forecasting & Planning” will be great opportunities to learn about the tools companies use to forecast and allocate resources. At Urbint, we have seen these processes become increasingly difficult as weather becomes more extreme and more of the industry’s workforce retires.

Another interesting event will be “Data Analytics and Reporting: Transforming Data Into Actionable Intelligence.” Whether or not data is the new oil is still up for debate, it’s clear the industries responsible for producing oil (gas companies included) are increasingly adopting new technology to help track operations, improve workflow and business processes — all in an effort to maintain the foundations of power.

These questions and topics are highly relevant to my day-to-day. As a deployment strategist at Urbint, I focus on understanding how gas operators can leverage AI to make more objective and smarter tradeoffs between public safety, reliability, and affordability.

Stop by our booth for some summer essentials — including sunglasses and water — to prepare you for that famous San Antonio sun, and learn how Urbint enables gas operators to make communities safer and more resilient.

Ruby Maa is a Deployment Strategist at Urbint. Ruby holds a degree in Chemical Engineering from the Massachusetts Institute of Technology (MIT). She previously worked in the healthcare technology industry and separately worked as a consultant to help large companies through their data strategy and digital transformation process.

Plugging into Industry Power at SGA’s Natural Gas Connect was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

Cities around the world have extensive systems of infrastructure, which play a vital role in a community’s capacity to thrive, grow, and achieve economic success. Providing electricity, gas, water, and connectivity underpin the very survival of a city’s inhabitants and businesses. Today, with roughly half the world’s population living in cities, increased urbanization is impacting the use and development of infrastructure on an unprecedented level.

Paris, the most populous city in France, is faced with a particularly challenging infrastructure environment. According to Eurostat, Paris has the highest urban density in Europe with a population of over 2.2 million — nearly 20,000 inhabitants per square kilometer. Besides serving the needs of a growing number of permanent citizens, the city’s gas and electric operators must also support 35 million annual tourist visits. To add another layer of complexity, the city manages over 10,000 building permits on an annual basis. These specific, entrenched challenges make the issue of infrastructure maintenance and modernization an absolute imperative in Paris.

A pressing infrastructure issue in Paris today is natural gas. Gas main replacement projects and other gas network repairs are more complex in a city where major excavation work can have an immediate effect on pedestrian and vehicle transit, impacting citizens, businesses, and tourists alike.

At Urbint, we analyzed the potential impact of natural forces and the built world on Paris’ gas distribution system to expose some of the city’s most unique infrastructure risks rooted from co-located infrastructure, weather and soil conditions, as well as construction activity. Our model captures all of the external factors that create infrastructure risk and applies AI to uncover the patterns and series of events that lead to specific incidences, such as methane leaks. This model of the world is dynamic and maintained in real time in order to anticipate threats from a changing world. Here’s a snapshot.

INTERESTED IN LEARNING MORE ABOUT OUR MODEL OF THE WORLD? LET’S TALK.

Collocation Index

Urbint measures the density of co-located underground infrastructure assets such as gas and water pipes, sewer mains, electrical and telecommunications cables. This index produces a score to better understand potential damage risk caused by excavation. The higher the collocation index score, the more maintenance, replacement, and consistent repair work are expected for a given area.

For Paris, the collocation index identified Châtelet–Les Halles as having the highest damage risk according to Urbint’s model. Châtelet–Les Halles is a major transit hub and shopping corridor in the city, which makes it a high-consequence area where the impact of excavation malpractice could prove catastrophic. Any dig performed in the area should be carried out with unflinching abidance to excavation best practices norms.

Weather and Soil Conditions

Paris weather tends to be moderately wet throughout the year, comprised of mild summers and winters, with few drastic temperature changes. However, recent weather volatility has affected temperature and precipitation patterns. In the face of increasing inclement weather, Urbint continuously monitors for climate extremes and trends that could impact our customers’ assets.

In January 2018, France experienced its heaviest rainfall in 50 years, with the national weather service reporting double the normal amount of rain nationwide. During the last week of January, the Seine River reached a maximum peak of approximately six meters, causing one of the worst floods in recent memory.

Almost a fifth of the Parisian gas network is located in the Seine’s 1-in-100 year floodplain. With increased volatility in precipitation, these assets could be increasingly exposed to stagnant waters and shifting soil conditions which lead to higher corrosion and damage risk.

During the colder months like January, gas assets are also exposed to higher rates of sustained freeze events. A sustained freeze event is a climatic condition where maximum apparent temperature does not exceed zero degrees Celsius (32 degrees Fahrenheit) during a given period. Sustained freezes can provoke sharp and unforeseen increases in odor call reports.

With that particular phenomenon in mind, Urbint’s software continuously computes unique slices of data like sustained freeze patterns that empower our AI models to learn more effectively about the world.

Sustained freezes is a leading indicator in projected leaks call frequency. These climatic events can be forecasted and empower utilities to preemptively staff their teams to respond to a surge in emergency response calls.

Urbint also looked at soil composition and identified that 40% of Parisian soils have features that are susceptible to corrosion risk as shown in the image below. Climatic conditions, like sustained freezes, don’t occur in a vacuum. For example, sandy soils can attenuate corrosion risks, while frost-prone soil types accentuate the effect of freeze-thaw events. Urbint is constantly on the lookout for a combination of factors that may compound aggravating effects on our customers’ assets.

Construction Activity

Rapid urbanization means most major metropolitan areas are experiencing high rates of development and construction activity. Urbint Lens for Damage Prevention specifically monitors and accounts for excavation work that puts an operator’s underground assets at risk.

Note: cities around the world have experienced steady increases in excavation activity over the past twenty years due to urbanization, modernization, and more.

To dive deeper into this phenomenon, Urbint looked at 20 years of Paris construction activity. By filtering demolition and new construction activity only, Urbint identified that over 35% of the Paris gas network is in areas with high historical rates of total or partial demolition (in red above). The most concentrated areas are in the city center, which happens to be the nucleus for tourism and commercial activity.

WANT TO KNOW IF YOUR CITY OR STATE MAKES THE GRADE ON EXCAVATION DAMAGE? CHECK OUT OUR URBINT REPORT CARD.

Why it Matters?

The root cause of risk for gas operators exists outside of their infrastructure. By examining hundreds of categories of external risk parameters that impact infrastructure, Urbint’s model of the world provides uncompromising situational awareness on an operator’s service territory. Drivers of asset degradation like corrosion and excavation hits are contingent on these external factors. Our model allows us to create unique and tailored digital representations of our customers’ service territory to help them better understand sources of risk.

The factors described above for the city of Paris are a sample of the hundreds of external risk factors that we monitor across the United States and selected regions worldwide.

Urbint AI software solutions mitigate risk and support confident decision making, ensuring reliable infrastructure for safer and more resilient communities. Learn more about how Urbint’s model of the world can help you ensure safe and reliable operations by reaching out to hello@urbint.com.

Sébastien Burgess is a Deployment Strategist at Urbint. Sébastien holds a degree in Conservation & Resource Studies from UC Berkeley and started his career in Mexico City focusing on urban planning challenges. He is a native of Paris and is passionate about climate change adaptation, urban resiliency, GIS, and energy policy issues.

Gas Operations in Paris: A Risk Assessment was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

The Urbint Report Card series provides a critical assessment of various aspects of the natural gas distribution landscape — from excavation damage to vintage pipe materials to leaks — and generates a performance grade for each area.

In the first part of our series, we’re analyzing excavation damage, both nationwide and state-by-state. Through this analysis, we established key metrics measuring damage prevention performance and also evaluated how states compare against each other in 2018. Read on to find out how your state(s) stacks up.

If you would like to know how your company scores, contact us at reportcard@urbint.com for a custom complimentary report.

Metrics Used

Our damage prevention grading leverages two metrics to assess performance:

1. Hit Rate
2. Excavation Damage per Exposed Facilities

Hit rate, specifically the number of excavation damages per 1,000 tickets, is a widely used metric in the industry to track damage prevention performance. It does a great job capturing improvements in call-in procedures. However, because hit rate normalizes the number of damages against the number of tickets, an increase in ticket count can overwhelm the metric and drive the hit rate down (regardless of actual damage count changes). In fact, from 2010 to 2018, approximately 120 gas operators saw a reduction in their hit rate, but an increase in their actual damage count, suggesting that hit rate does not provide the complete picture. We believe that, while hit rate is a widely used metric in the industry, it should be paired with a metric that does not use the number of tickets as the denominator.

As a result, absent the collection of additional data, we utilized a separate metric to measure performance: excavation damage normalized by service territory size. This metric normalizes the number of excavation damage against the size of a territory in terms of main miles and services. Because service territory changes are less volatile, we believe that this provides a more accurate assessment of year-over-year changes and allows us to compare states that might have very different 811 call behaviors.

Based on these two metrics, we’ve established a grading system to measure damage prevention improvements nationwide and for individual states. For further details on the grading system, please refer to the end of this post.

Nationwide

Hit Rate Improvement Score (2018): C

The chart below shows how the nation’s hit rate has changed since 2010.

In 2018, the industry saw a statistically insignificant reduction in hit rate. The hit rate went down from 2.75 damages per 1,000 tickets in 2017 to 2.73 in 2018. With a hit rate that effectively remained the same between 2017 and 2018, the nation received an average grade of “C” for hit rate improvement.

It is worth observing a cyclical pattern in hit rate changes, with some years seeing significant reductions followed by a few years of little to no change. The most recent significant reduction in hit rate was from 2016 to 2017, when the nation reduced the hit rate from 3.03 to 2.75.

Damage per Service Territory Size Improvement Score (2018): D

The chart above shows how the nation’s damage per territory size has changed since 2010.

In 2018, the industry saw a slight increase in excavation damages per service territory size. This notable increase in damages per territory size earned the nation a “D” grade.

A historical review of this metric shows significant changes year-to-year. 2016 saw a major increase in damage per territory size, followed by a reduction in 2017.

2018 State Comparison

Hit Rate

Top 10 states with the lowest hit rate in 2018:

1. Maryland, with a hit rate of 0.95
2. Virginia, with a hit rate of 0.96
3. Arizona, with a hit rate of 1.12
4. Connecticut, with a hit rate of 1.27
5. New York, with a hit rate of 1.62
6. Nebraska, with a hit rate of 1.67
7. Wisconsin, with a hit rate of 1.68
8. Indiana, with a hit rate of 1.71
9. Washington D.C., with a hit rate of 1.74
10. New Jersey, with a hit rate of 1.91

Top 10 states with the highest hit rate in 2018:

1. Arkansas, with a hit rate of 6.8
2. Idaho, with a hit rate of 6.38
3. Alaska, with a hit rate of 6.14
4. Michigan, with a hit rate of 4.95
5. South Dakota, with a hit rate of 4.80
6. Mississippi, with a hit rate of 4.73
7. Alabama, with a hit rate of 4.23
8. Tennessee, with a hit rate of 4.06
9. Texas, with a hit rate of 4.02
10. New Mexico, with a hit rate of 3.62

Excavation Damage Per Service Territory Size (2018)

Top 10 states with the lowest damages per size in 2018:

1. Rhode Island, with a score of 2.58
2. Arizona, with a score of 2.72
3. New York, with a score of 2.81
4. West Virginia, with a score of 2.84
5. Connecticut, with a score of 2.92
6. New Mexico, with a score of 3.02
7. Nevada, with a score of 3.24
8. Wisconsin, with a score of 3.4
9. Vermont, with a score of 3.58
10. Massachusetts, with a score of 3.59

Top 10 states with the highest damages per size in 2018:

1. Florida, with a score of 12.65
2. Alabama, with a score of 12.6
3. South Carolina, with a score of 11.7
4. Arkansas, with a score of 10.16
5. Texas, with a score of 9.91
6. Idaho, with a score of 9.90
7. North Carolina, with a score of 9.10
8. Mississippi, with a score of 8.95
9. South Dakota, with a score of 8.75
10. Tennessee, with a score of 8.44

Finding the Right Solutions to Succeed

While a few territories are seeing straight-As, improvements are still needed across the board. The reality is that increasingly complex operating environments — along with construction and urbanization leading to higher prevalence of issues like no-calls and mislocates — make damage prevention a continuous area of focus for LDCs. Interested in learning more about how artificial intelligence can enable you to minimize excavation damage? Check out Urbint Lens for Damage Prevention.

Analysis Details

About the Scoring Methodology

The hit rate was calculated by dividing the number of excavation damages by 1,000 tickets for the territory. Nationwide and statewide aggregation involve aggregating all damages in the nation or state, and then dividing it by the aggregated ticket count.

The excavation damages per territory size involved dividing the number of excavation damages by a Urbint-calculated service territory size metric. The service territory size metric accounts for the number of main miles and the number of services, balancing both as equally important.

From there, below are the grade thresholds used.

For hit rate (damages per 1,000 tickets):

• A: Less than 1.8
• B: Between 1.8 and 2.97
• C: Between 2.97 and 4.81
• D: Between 4.81 and 9.40
• F: Greater than 9.40

For damages per territory size:

• A: Less than 2.91
• B: Between 2.91 and 4.73
• C: Between 4.73 and 6.95
• D: Between 6.95 and 11.07
• F: Greater than 11.07

For each metric, these thresholds are established by taking the 20–40–60–80 percentile across the 2010 to 2018 data.

To score based on improvement rate, the percentage change was first calculated for both the hit rate and the damage per territory size. The percentage change is determined by taking the difference and dividing that by the previous year’s metric. As an example, the hit rate percentage change for 2018 would be taking the change in hit rate from 2017 to 2018, and dividing it by the hit rate in 2017.

From there, the percentage change is scored based on the following ranges:

• A: Greater than a 5% decrease
• B: 1–5% decrease
• C: 1% decrease to 1% increase
• D: 1–5% increase
• F: Greater than a 5% increase

Source and Disclaimers

Our analysis leverages data from the Pipeline and Hazardous Materials Safety Administration (also known as “PHMSA”) Annual Reports, officially known as “Annual Report Data from Gas Distribution, Gas Gathering, Gas Transmission, Hazardous Liquids, Liquefied Natural Gas (LNG), and Underground Natural Gas Storage (UNGS) Facility Operators”. The original source of the data is available here.

Because PHMSA only started tracking what commodities each operator was distributing in 2015, and because our analysis goes back to 2010, before specific commodities were being tracked, it does include data for other types of gas.

That said, 2018 data shows that Natural Gas is the predominant commodity represented in this dataset for the nation and also for the majority of the states. In 2018, Natural Gas represents 99.8% of main miles reported and 99.8% of the services reported. For all states, with the exception of Hawaii and Puerto Rico, Natural Gas represents at least 95% of main miles reported and at least 95% of services reported. Hawaii and Puerto Rico are excluded from the analysis because their main miles and services do not primarily serve Natural Gas. Please inquire for the exact percentage for your state specifically.

The Urbint Report Card Series: Stacking Up Excavation Damage Across the States was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

This piece originally appeared in Utility Dive.

The following is a contributed article by Brien Sheahan, Commissioner of the Illinois Commerce Commission, Jimmie Zhang, a legal and policy advisor to Commissioner Sheahan, and Corey Capasso, CEO of Urbint.

We live in the age of big data.

Utilities used to read customer meters 12 times per year in person. Now smart meters generate reports every 15 minutes remotely. That’s thousands of times more information at a fraction of the cost. In combination with an ever-growing database of environmental data external to the utility, utilities receive more data than they know what to do with.

According to a survey from Utility Analytics Institute, less than half of utilities have standardized tools and procedures to leverage such data. That shouldn’t be surprising, the large influx of data created in the last decade have rendered traditional analytical techniques relatively impotent.

DA systems have become an essential tool for quickly locating and isolating grid faults, but utilities often struggle with high network latency. Learn how to improve the responsiveness and reliability of your systems.

The utility industry isn’t alone in facing this overwhelming data challenge. The response has been to utilize innovative solutions like artificial intelligence to bring order and generate predictive value behind such data.

The success of such efforts is evident across many industries.

MIT showed that their artificially intelligent model can predict the onset of breast cancer 5 years in advance. Amazon now shows what you want to buy before you realize it. The trends that the utility industry faces today necessitates the consideration of such technology.

New infrastructure challenges

In the last half-century, trends in urbanization have resulted in over 80% of the U.S. population living in urban areas, up from 64% in 1950. This dramatic uptake in population density has engendered a series of new challenges for infrastructure development and operation.

The big question is, how do we accommodate this growth and increasing complexity while ensuring the sustainability and resilience of our infrastructure?

The current state of our communities is increasingly at risk, considering the growing stress placed on our aging utility systems and the human capital necessary to maintain and operate them.

Nearly half of all “billion-dollar disasters” have occurred in the last 10 years (adjusted for inflation). The utility workforce is aging: More than 25% of the current workforce is expected to turn over in the next 5–7 years. And, as regulators and infrastructure owners and operators understand, our aging system poses a severe risk to public safety.

Just last month in Durham, a gas explosion killed one and injured 25 others. Similar incidents involving aging cast iron mains in cities such as Allentown, Austin and Philadelphia have killed dozens of people and injured many more.

The Merrimack Valley gas explosion in Massachusetts last September even garnered swift attention from federal lawmakers, leading to the introduction of the Leonel Rondon Pipeline Safety Act in the Senate to tighten pipeline safety regulation.

Unsurprisingly, local distribution companies (LDCs) face overwhelming operational challenges to ensure public safety while minimizing rate impact. Over $270 billion dollars are being poured into pipeline upgrades across the country to increase reliability and reduce methane leaks.

At the same time, policy makers and state regulatory commissions are continuing to raise expectations for utility performance as a result of public safety and environmental concerns on gas distribution systems.

One prominent issue that LDCs are grappling with is how to better understand the external risks posed to their system.

For example, a major impact of urbanization is that it precipitates a host of new construction and redevelopment activity. As a result, gas distribution operators are observing an increase in excavations that could potentially damage pipelines and cause severe hazards to the local community and environment.

Using AI

Illinois has been looking at ways to reduce the number of third-party damages. These damages can be caused by error from numerous sources: excavators, contractors “locating” LDC gas lines, or instances where someone digs but doesn’t call the customary “call before you dig” number.

Enabled by a forward-looking regulatory environment, one Illinois LDC has been tackling these types of challenges by working with firms incorporating the latest developments in AI.

By modeling the physical world that represents external threats facing our infrastructure, and applying AI, utilities can now predict instances of a variety of hazards allowing them to make smart operational decisions.

Methane leaks, areas of corrosion, locations of cross-bores and high volumes of odor calls can all be predicted with AI.

Third-party damages — when a construction company nicks a gas line, for example — are becoming more frequent in urban areas. AI is helping utilities intervene and potentially prevent 25% or more of potential damages caused by excavation.

States like California have formally recognized the benefit of using technologies such as predictive analytics as best practices in detecting and preventing harm.

Other areas of the utility sector are also leveraging technology advancements in AI. The California Wildfires last summer have focused concern for utilities around the country to develop strategies to mitigate such risks.

The primary cause of wildfires is the contact of energized transmission and distribution wires with vegetation susceptible to fire. While existing rules and methods on vegetation management and clearance areas already exist, instead of blanketly increasing resources for tree-trimming and other forms of vegetation management, utilities in California and Colorado are proactively using drone-mounted LiDAR technology and corresponding AI software to conduct rigorous assessments of overhead electric systems.

The use of such technology allows novel ways of identifying at-risk infrastructure and the development of more cost-effective replacement strategies.

Predictive technologies are readily available to identify growing risks and dramatically reduce the rate and consequence of incidents, yet the uptake by the utility industry as a whole has been sluggish. Because the basic framework for utility regulation has not evolved from its inception at the turn of the 20th century, state regulators bear some responsibility for the slow adoption of new technology.

The current conservative cost-of-service framework does not incentivize utilities to experiment or adopt innovative solutions. But increasingly innovation is necessary to provide the services and reliability that customers want and to ensure public safety.

Some states are moving in the right direction by removing the disincentive for investing in innovation, such as considering the capitalization of cloud computing expenditures.

The accelerating pace of new technology, in combination with the recent catastrophic failure of basic utility infrastructure around the country, require policymakers and regulators to consider creating a more attractive environment for investment and innovation.

Close collaboration among private sector innovators, utilities and state public utility commissions will be necessary to identify critical areas where new technologies can keep communities safe and prevent potential life and property threatening events. One potential framework to enable such collaboration is the idea of a regulatory sandbox that has been gaining traction in the finance industry and starting to be considered by energy regulators.

While it has dependably supported the economic, industrial and consumer needs of the last century, the evolutionary pace of utility operations is no longer acceptable in an environment of accelerating urban transformation.

When it comes to solving modern challenges, we cannot rely on yesterday’s regulatory frameworks and technological solutions. A focus on innovation is paramount to address challenges of today and the future.

Averting Disasters with Innovation: Utilities’ Pressing Need for Predictive Technology was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

The stakes of reliable infrastructure have never been higher. Factors like aging infrastructure, urbanization, drastic changes in weather patterns and workforce turnover mean that gas utilities are being faced with unprecedented complexities. At Urbint, we’re focused on empowering those operators to make more effective decisions in a challenging operating environment using artificial intelligence.

When we walked into our first meeting with a major utility company several years ago, the premise was not what Urbint has become today. What started out as a conversation about data turned into the realization that our technology could actually help solve the biggest, most complex challenges facing utilities today.

Flash forward to 2019. With over 80% of Americans living in urban areas and the cost of infrastructure replacement climbing toward over $4.5 trillion by 2025, the tough calls operators have to make every day on issues like methane leak inspections, damage prevention, and pipe replacement prioritization affect millions of citizens.

Balancing risk mitigation and resource allocation in a changing world has never been more important in building and maintaining resiliency in our communities. It became clear to us early on in our journey that AI was the only way to meet both existing challenges as well as ones never seen before because it has the power to identify these patterns of events and factors that have historically been difficult to account for.

Our investors agreed with Urbint’s mission of making communities safer and more resilient through AI, and capitalized the business with $18.5 million in funding from investors including Energy Impact Partners, Urban US, Fifth Wall, and a handful of strategic angel investors. These firms not only share Urbint’s vision for safer, more resilient communities, they also specialize in our space: shaping the energy landscape of the future, bringing technology to the built world, and reimagining how communities can thrive.

With this capital and proven success enabling some of the largest gas utility companies in the world to maximize risk reduction for every dollar spent, we’re continuing to scale. The addition of Amen Ra Mishariki as Head of Urban Intelligence and a fast-expanding best-in-class team, along with the addition of new gas utility partners, means that our headquarters in New York City is becoming the center point of an evolution in infrastructure safety and resiliency. We’ve started in the gas sector and will expand in the near future into other areas such as electric and municipalities.

Despite the unprecedented challenges facing communities around the globe today, our outlook is positive. We know that the challenges of aging infrastructure can’t be solved with aging technology, and we’re focused on building solutions that solve these problems in new ways — delivering cutting-edge technology to infrastructure operators and utilities without disrupting the established workflow, enabling them to deploy the right intervention at the right time in the right place.

We believe that the technology, team and vision we’re building at Urbint today will be the foundation of a new way of thinking about infrastructure — one that prepares us for the kind of complexities and challenges we may not even be able to imagine today. Anticipating the future is exactly what we’re working to do, and the greatest proof of our success will be safer, more resilient communities supported by the most reliable infrastructure possible.

The Future of Urbint: How We’re Transforming Gas Infrastructure Safety with AI was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

How do you know what your organization is good at? Start with your culture. When and where do people eat lunch? Are meetings starting and ending on time? What kind of product development process is used? How does the product roadmap get made? The accumulation of these quirks, the expression of the history and values of a company, are the culture of the company. Any work that aligns with the culture will naturally be a little easier, turn out a little better, then work at odds with your organizational tendencies. Put another way, the culture of your company represents what you’re innately good at. Great companies turn their natural strengths into value.

Jeff Bezos famously built a culture where communication was shunned and teams had to feed themselves with two pizzas at a time. In fact, Bezos went as far as to mandate that teams should only communicate via API, no informal communication, no shared repositories of information, just API. This essentially turned the teams within Amazon directly into microservices. In turn this has allowed Amazon to externally monetize every major service that is developed within the company without much additional effort. If it works at internal Amazon scale, it’s likely to work at external Amazon scale. Fifteen years later everyone is talking about how two pizzas led to massive successes like AWS and Amazon Marketplace.

Steve Jobs is another CEO famous for having a massive impact on company culture. In Jobs’ case, he is famous for going over every detail, demanding excellence, and crucially putting the customer first. In 1997, Jobs’ said, “You’ve got to start with the customer experience and work backwards to the technology…” This wasn’t just rhetoric — his outbursts led a lot of very smart people over the years to work hard to solve these central mandates. This spirit continues today; in one story, Bob Messerschmidt had to come up with an entirely new heart rate sensor technology so that consumers could stay “on trend”. Obsessively putting the consumer first has allowed Apple to build the most valuable company in the world with a startlingly small array of products.

The takeaway from these famous success stories isn’t, “turn your teams into microservices”, or “empower a central group or figure to demand excellence on behalf of customers”. You can’t build a successful company by backfitting the unique things other companies do into recipes for success. One of my favorite business books, The Halo Effect, lays out this idea in detail. What you can do, is understand what unique aspects of your culture influence the way you do business, and align your business strategy to them. Like the Joker, find what you’re good at, and turn it into value.

At Urbint, one of our core values is being dialectic. I’ve had long, serious discussions about whether we should refer to parts of our graph as vertices (obviously) or nodes. The core of our product has always been in trying to distill the true facts about a city from all the myriad data sources available. The business really found traction when we aligned with a business problem, making predictions via machine learning, where the truthfulness of data dictates the quality of results. We’re aligned to a space where our cultural tendency to pursue the truth leads directly to business value.

What Heath Ledger’s Joker can Teach About Business Strategy was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

Binary classification with strong class imbalance can be found in many real-world classification problems. From trying to predict events such as network intrusion and bank fraud to a patient’s medical diagnosis, the goal in these cases is to be able to identify instances of the minority class — that is, the class that is underrepresented in the dataset. This, of course, presents a big challenge as most predictive models tend to ignore the more critical minority class while deceptively giving high accuracy results by favoring the majority class.

Several techniques have been used to get around the problem of class imbalance, including different sampling methods and modeling algorithms. Examples of sampling methods include adding data samples to the minority class by either duplicating the data or generating synthetic minority samples (oversampling), or randomly removing majority class data to produce a more balanced data distribution (undersampling).

Different algorithms are also more suitable for class imbalance problems, including those based on boosting, such as Adaptive Boosting (AdaBoost). AdaBoost works to improve the performance of what are known as weak learners (poor predictive models, but better than random guessing). AdaBoost iteratively builds an ensemble of weak learners by adjusting the weights of misclassified data during each iteration. For the training of the first weak learner, AdaBoost assigns equal weight to each training set sample. For each subsequent weak learner, the weights are recalculated such that a higher weight is assigned to samples that the current weak learner misclassified. This weight determines the probability that the sample will appear in the training of the next weak learner. For this reason, boosting algorithms like AdaBoost are particularly useful for class imbalance problems because higher weight is given to the minority class at each successive iteration as data from this class is often misclassified.

Data sampling methods combined with boosting can be an effective way to deal with class imbalance problems. To demonstrate, let’s use Lending Club loan data (available here) to try to predict whether someone will default on a loan using two interesting sampling approaches that have been suggested and combined with AdaBoost: SMOTEBoost and RUSBoost.

SMOTEBoost is an oversampling method based on the SMOTE algorithm (Synthetic Minority Oversampling Technique). SMOTE uses k-nearest neighbors to create synthetic examples of the minority class. SMOTEBoost then injects the SMOTE method at each boosting iteration. The advantage of this approach is that while standard boosting gives equal weights to all misclassified data, SMOTE gives more examples of the minority class at each boosting step. Similarly, RUSBoost achieves the same goal by performing random undersampling (RUS) at each boosting iteration instead of SMOTE.

Trying it out: data preparation

First, I import the data and get rid of all entries that do not correspond to loan_data = “Fully Paid” or “Charged Off”. I set the “Fully Paid” target to 0 and “Charged Off” to 1.

import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.ensemble import AdaBoostClassifier
from sklearn.metrics import classification_report
from sklearn.metrics import average_precision_score
from sklearn.metrics import precision_recall_curve
from sklearn.utils import resample
from imblearn.over_sampling import SMOTE
import smote
import rus

# read in data and set target variable

df = pd.read_csv("lending-club-loan-data/loan.csv", 
                 low_memory=False)
df['target'] = ''
df.loc[df.loan_status == 'Fully Paid', 'target'] = 0
df.loc[df.loan_status == 'Charged Off', 'target'] = 1
df = df[df.target != '']

I use the following loan features to build the binary classifier:

• delinq_2yrs: number of 30+ days past-due incidences of delinquency in the borrower’s credit file for the past 2 years
• dti: ratio calculated using the borrower’s total monthly debt payments on the total debt obligations, excluding mortgage and the requested LC loan, divided by the borrower’s self-reported monthly income
• home_ownership: home ownership status provided by the borrower during registration. Values are RENT, OWN, MORTGAGE, OTHER
• grade: LC assigned loan grade
• int_rate: interest rate on the loan
• purpose: a category provided by the borrower for the loan request
• revol_util: revolving line utilization rate, or the amount of credit the borrower is using relative to all available revolving credit
• total_rec_late_fee: late fees received to date

features = ['grade','home_ownership','dti','purpose','int_rate',
            'delinq_2yrs','revol_util','total_rec_late_fee',
            'target']
df_model = df[features]

After removing any rows with missing data, the default loans represent about 18% of the chosen dataset.

Finally, before resampling and training the classifier, I transform the categorical data into numeric representation with one hot encoding and split the data into training and testing sets. Sampling methods are only used on the training set while the testing set maintains the same class imbalance seen in the original dataset.

X = pd.get_dummies(df_model.drop('target', axis=1))
y = df_model['target'].tolist() 
X_train, X_test, y_train, y_test = train_test_split(X, y, 
                                                    test_size=0.3, 
                                                    random_state=42)

Initial Results

I first train a few models with:

1. just AdaBoost,
2. SMOTE sampling and then AdaBoost,
3. and RUS sampling and then AdaBoost

def adaboost(X_train, X_test, y_train):
    model = AdaBoostClassifier(n_estimators=100, random_state=42)
    model.fit(X_train,y_train)
    y_pred = model.predict(X_test) 
    return y_pred

# AdaBoost
y_baseline = adaboost(X_train, X_test, y_train)

# SMOTE
sm = SMOTE(random_state=42)
X_train_sm, y_train_sm = sm.fit_sample(X_train, y_train)
y_smote = adaboost(X_train_sm, X_test, y_train_sm)

# RUS
X_full = X_train.copy()
X_full['target'] = y_train
X_maj = X_full[X_full.target==0]
X_min = X_full[X_full.target==1]
X_maj_rus = resample(X_maj,replace=False,n_samples=len(X_min),random_state=44)
X_rus = pd.concat([X_maj_rus, X_min])
X_train_rus = X_rus.drop(['target'], axis=1)
y_train_rus = X_rus.target
y_rus = adaboost(X_train_rus, X_test, y_train_rus)

Results from these implementations provide an idea of any improvement gained from using SMOTEBoost and RUSBoost.

Sampling has significantly improved the recall of the minority class labeled “Default”, with the largest improvement seen from using RUS. Note that the number of samples generated or removed in this implementation is such that the minority and majority classes have equal number of samples before running AdaBoost.

Setting up SMOTEBoost and RUSBoost

For SMOTEBoost implementation, I use the scikit-learn AdaBoost implementation with a modified fit function found here. Before each boosting step, a SMOTE resampling calculates new synthetic examples for the minority class. I use the default k=5as the number of nearest neighbors to calculate the new data and set n_samples=300, which is the amount of synthetically generated examples to be used in each boosting step. Care should be taken with setting the n_samplesparameter such that the number of minority class samples does not become greater than the majority class samples, causing an opposite class imbalance. The weights given to the new data is normalized based the current boosting step’s training set.

Similarly, RUSBoost implementation is a modified fit function (found here) where before each boosting step the majority class is undersampled. I set n_samples=300 as the number of samples to remove in each boosting step.

target_names = ['Not Default', 'Default']
for algorithm in [smote.SMOTEBoost(n_estimators=100, n_samples=300),
                  rus.RUSBoost(n_estimators=100, n_samples=300)]:
    algorithm.fit(X_train, y_train)
    y_pred = algorithm.predict(X_test)
    print()
    print(str(algorithm))
    print()
    print(classification_report(y_test, y_pred, 
                                target_names=target_names))

The full implementation of this loan default example can be found here.

Assessing results

For problems with class imbalance, metrics such as precision, recall, and f1-score give good insight to how a classifier performs with respect to the minority class. Depending on the problem, the goal is to optimize precision and/or recall of the classifier. In this case, I want a model that catches the most number of instances of the minority class, even if it increases the number of false positives. A classifier with a high recall score will give the greatest number of potential loan defaults, or at least raise a flag on the most vulnerable cases. In this example, RUSBoost slightly outperforms SMOTEBoost when looking at the recall of the minority class for the given setup.

SMOTEBoost:

RUSBoost:

So, how does this compare to just using AdaBoost without sampling, or using each sampling method just once before running AdaBoost? The results of each implementation for the minority class are summarized below.

The best sampling method for this modeling setup is RUS. It, by far, gives the highest recall for the minority class.

Other useful methods to evaluate a classifier with class imbalance are ROC and precision-recall (PR) curves. Since I’m not as interested in how the model performs on the negative class (‘Not Default’ class), I plot just the precision-recall curves, which do not consider true negatives. These PR curves can be used to compare RUS to RUSBoost results.

Conclusion

While SMOTEBoost holds promise of improving AdaBoost on unbalanced data by broadening the scope of the minority class, in this case, it showed to be the weaker of all sampling approaches for the given model setup. Similarly, RUSBoost did not perform as well as simply RUS sampling followed by AdaBoost.

In general, both sampling methods have their drawbacks, such as the increased model training time for any method that generates additional data, and loss of valuable data for any undersampling technique. Also, when approaching any classification problem with class imbalance, or any machine learning problem altogether, proper data preparation is critical, including making sure that the quality of data is sound as ensemble methods attempt to correct any misclassification. There is a lot of room for improvement in massaging the loan data used in this example to make sure it is getting the most out of the different sampling techniques. The same can be said about implementing the different sampling boosting methods. Some approaches are more fitting for specific types of problems and the degree of imbalance in the data.

At Urbint, we help urban operators solve complex infrastructure problems by producing actionable insights and predicting often highly critical events. In our work, class imbalance is a frequent occurrence. Most events we are trying to predict are rare compared to the negative class. Sampling and boosting are just some of the techniques we employ to optimize the performance of our models.

Using SMOTEBoost and RUSBoost to deal with class imbalance was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

After spending a while struggling to build my own dotfiles from scratch, the team discovered an amazing package of configuration called Doom (a huge thank you to Henrik Lissner for putting it together). For any Emacs purists reading this, you can probably close the tab in disgust now. Doom is very much built around Evil Mode and is meant to make VI users feel at home. For anybody looking to really get started being productive with Emacs, I thought it would be helpful to compile the most common functions that I use, and how they are configured in Doom. If you are an Emacs pro and just want to learn about Doom, some of this will be repetitive, and you would probably be better served just to read the README on the Doom project, but if you are looking to started and still feel a bit lost in Emacs, hopefully this will help you out.

To get started, make sure you have Emacs installed and then follow the Doom installation guide: https://github.com/hlissner/.emacs.d#installation

Emacs Key Binding Philosophy

The three core concepts for Emacs key bindings are modifiers, chords and sequences.

A modifier key is any one of the following. Anything else is a “non-modifer key”:

A chord is made up of zero or more modifier keys pressed at the same time as a single non-modifier key. For example, “C-x” is a chord for pressing Ctrl and “x” at the same time and then releasing them. “C-s-v” is a chord for pressing Ctrl, Command (or Windows), and v at the same time.

A sequence is a series of chords, pressed and released in a row. The most important command in Emacs, “C-x C-c”, is a great example of a sequence. You press “Ctrl” and “x” at the same time, then release them, then press “Ctrl” and “c” at the same time and then release them and voila, you are in the process of exiting Emacs.

Doom Note: Emacs has a special extension called “evil mode” that emulates a lot of vi like functionality. Doom is very strongly centered around evil mode, and a lot of the rest of this article involves using the various key bindings that are configured with it. The biggest piece is the concept of an “evil leader” which is basically a key that you press as the first step in a sequence that then opens up a new branch of possible commands. In Doom, the evil leader is set to the Space Bar by default. So, for example, to split the screen into two side-by-side windows, you would press “SPC w v”. Keep this concept of the leader key in mind as we go through practical examples.

Setting Up Projects

If you are more familiar with IDEs like Eclipse or IntelliJ, you probably already have a concept of a project in your mind. It’s basically a folder for a particular codebase (probably under version control) that is pretty much a cohesive unit that you work on independently. In Emacs, this grouping and identification is usually managed by Projectile. With Doom, this is installed by default.

In order to make use of this functionality with the projects that you already have, you need to let Projectile know where the projects reside. Inside of Doom, you do this by modifying the init.el inside your own custom configuration folder:

~/.emacs.d/modules/private/<your-user-name>/init.el

For example, my username is “jdemaris” so my folder is ~/.emacs.d/modules/private/jdemaris is my custom folder. Keeping all of your changes in here makes it safer to update the Doom config whenever new versions come out. Inside of my init.el file, I have added a number of Projectile projects:

(projectile-add-known-project “~/Projects/playground/elixir”)
(projectile-add-known-project “~/Projects/playground/otp”)
(projectile-add-known-project “~/Projects/playground/expostal”)
(projectile-add-known-project “~/Projects/playground/benchfella”)

You should be able to make the folder and add an init.el file for yourself as well, pointing to one or two projects that you would like to be editing with Emacs.

Opening & Navigating Projects

Once you have this in place, open up Emacs and we can try switching to the project! Once emacs is open, try executing the following sequence:

SPC p p

That is, press space bar, release it, press p, release it and then press p again and release it. You will see a small modal slide up from the bottom, listing out the projects you defined!

You can pick amongst them using the arrow keys, or (if you’re already familiar with some common key bindings) ctrl-j to move down and ctrl-k to move up. Better still, if you have a long list of projects to switch between, you can start typing the name to filter down the list! Once you’ve found the project that you want to work on and selected it, you can hit Enter.

This won’t open the project up directly, so don’t be surprised when you don’t see content up top yet. Instead, this will open the list of files in that project in that small modal in the bottom (called the “mini buffer”). Again, you can use the same navigation methods listed above to select the first file that you want to open. Once you’ve found the file and hit enter, you will see the file in the main window and you can get started editing on it.

By default, the screen will be in “normal” mode, which means that it is expecting commands (just like if you opened up vi). If you hit the “i” key, it will convert to INSERT mode and you can type text as you normally would in vi. To go back to normal mode, hit the Escape key. Once you’re back in normal mode, you can navigate to other files within the same project by using the sequence:

SPC SPC

This opens up the already familiar navigator and lets you pick another file to switch to within the same project. If you want to open a file from a different project, you can use the “SPC p p” sequence from earlier again. If you would like to switch between files that you have already opened (these currently open screens are called “buffers” in Emacs) then you can use:

SPC ,

Evil mode emacs is so vi-esque that your old friends “:w” and “:q” are available to save and quit as well.

Window Management and Navigation

What good is an editor if you can only have a single thing on screen at one time? Not much good, and it makes your 4k monitor cry to be so underutilized. In Emacs, you can easily split the screen horizontally and/or vertically into different windows that contain different data. In the background, Emacs has a bunch of buffers open, one for each file you have opened to edit. When you open a window, it assigns a buffer to that window. If you happen to have the same file open in two different windows, then you are actually editing the same buffer! So if you edit the contents of one window, you’d see that change happening on all of the other windows with that same file. This can actually be super useful if you have a long file and you want to look at one part of it while you edit the other part.

So how do you create these new windows? To split your current window into two side-by-side windows with Doom, press:

SPC w v

This will open up a split window for you like this with the same buffer open in both windows:

You may be starting to see a pattern in these key presses. They are actually categorized! SPC is the evil leader, so it indicates that you want to execute some command. “w” is actually for the “window” category. If you forget exactly which key to press, but you can remember the category, then Doom has the plugins set up to guide you through the process. For example, pressing:

SPC w

will open up this guidance section in the minibuffer:

You can see that “v” executes evil-window-vsplit. Can you see what key to press instead of “v” if you wanted to split the window into top and bottom? I’ll give you a hint — it’s just called evil-window-split.

Once you have some windows open, you need to be able to move in between them. Your mouse will actually work, but that’s not very Emacs-esque. There are a few different ways, but my favorites to move around are:

And what if you want to get rid of some of these windows? To close the currently selected window:

SPC w c

You can also use Ctrl-X 0 (zero). There are a ton of other things you can do with these windows. Go through the list of options shown after pressing “SPC w” and try them out!

File Structure Navigation (neotree)

What if you don’t quite know what your project layout looks like and you’re more comfortable seeing a tree of your file system? Fear not! Neotree is one of the most widely used file system tree views in Emacs, and it comes preconfigured with Doom.

To open the Neotree view:

SPC o n

You can do this from pretty much anywhere (except insert mode — so think anywhere you would be able to use :w or :q). The really great part is that this is actually a smart open! If you are not currently inside of a Projectile project, then it will start with your home folder. If, however, you already selected a project and are working in that context, it will open up the tree view of just that project.

You can use the arrow keys to move up and down on tree, or if you’re already used to it — h,j,k and l also work. Pressing ENTER on a file will open it up in the original window and move the focus of your cursor over to that window.

To create a new file in neotree so you can start editing it, press the “c” key. It will open up a dialog in the minibuffer to let you specify the path to the file. Once you are done typing the path, hit enter and it will close the minibuffer and create the file (and any parent folders) in neotree for you to select and open.

To delete a file or folder that you have selected in neotree, press:

C-c C-d

If you are currently in the neotree window and you want to close it, you can just press the Escape key or the “q” key. “SPC w c” will still work, but escape is a lot more intuitive to me since I feel like I am closing out a temporary dialog.

Using the terminal

Finally, what if you want to do something in the CLI that you don’t have (or don’t know) the keybinding for yet? Do you have to leave Emacs for this? Fear not!! You can open up a terminal instance right in Emacs as one of its windows. I personally find this handy for doing git-related commands since my muscle memory for git is much more tied to the CLI than to the editor.

To open up a CLI window, press:

SPC o t

You may have gotten the idea of those categories from earlier. Since we use “SPC o t” for the terminal and “SPC o n” for neotree, you can probably guess that “o” is the category for opening popup windows.

You can perform whatever CLI operations you want in here, and even use the window navigation keys to switch between your editing windows and the CLI window. Once you are done with the terminal, you can close it by pressing:

C-x 0 (zero)

This will close it back out and return you to your normal editing.

Conclusion

One final tip — if you are sure that a command must exist for something but you aren’t sure what it’s called, press Alt-X (or “M-x”) on your keyboard and you will get the buffer to actually type the named commands into Emacs. Each of the key sequences listed here is actually tied to essentially a function name, and you can call that function by searching for it as well. If you decide that you don’t actually want to call a function after all, you can always press Ctrl-G to cancel whatever you were in the middle of doing.

This is only barely scratching the surface of what kind of power you have available, but hopefully it can help you get over the initial learning curve and start down the path of Emacs wizardry!

Emacs Doom for Newbies was originally published in Urbint | Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.