David paused to let the video sink in before resuming. “We provide fully customized recommendations, because each person is motivated by specific kinds of language, styles of communication, and reasons. Let’s use another example. An employee is going to ask for an extended vacation; he’d like to make a compelling case for granting his request. What will motivate his manager? Should he mention he’s been working overtime, or that he needs to spend time with his kids? What if he’s planning to visit the Grand Canyon, a place his manager associates with good memories?
“The answer,” David went on, as he paced back and forth in the front of the room, “depends on the person you’re emailing. So ELOPe customizes its analysis not only to what the sender is asking for, but for what the recipient is motivated by.”
David noticed Rebecca Smith standing in the doorway listening to the presentation. In a sharp tailored suit, her reputation hovering about her like an invisible aura, the Avogadro CEO made for an imposing presence. Only her warm smile left a welcoming space in which an ordinary guy like David could stand.
She nodded to him as she came in and took her seat at the head of the table.
“What you’re describing,” Kenneth asked, “how does it work? The natural language processing ability of computers doesn’t even come close to being able to understand semantics. Have you had a miracle breakthrough?”
“At the heart of how this works is the field of recommendation algorithms,” David said. “Sean hired me, not because I knew anything about language analysis, but because I was a leading competitor in the Netflix Prize. Netflix recommends movies you’d liked to watch. The better they can do this, the more you as a customer enjoy using their service. Several years ago, Netflix offered a million-dollar award to anyone who could beat their own algorithm by ten percent.
“What’s amazing and even counterintuitive about recommendation algorithms is that they don’t depend on understanding anything about the movie. Netflix does not, for example, have a staff of people watching movies to categorize and rate them to find the latest sci-fi space action thriller I may like. Instead they rely on a technique called collaborative filtering, where they find other customers like me and analyze how those customers rated a given movie, to predict how I’ll rate it. Sean’s insight was that since natural language analysis struggles to understand semantics, it would be best to start with an approach that doesn’t rely on understanding, but instead one which utilizes patterns.”
When David received nods from the audience, he went on. “ELOPe parses billions of emails, comparing the language used and how the recipient reacted. Was the response positive or negative? Compiled over thousands of messages per person, and millions of people, we can find a cluster of users similar to the intended recipient of an email and analyze how they respond to variations of language and ideas to find the best way to present information and make compelling arguments.”
Now there were puzzled looks and half-raised hands as people around the room tried to ask questions. David forestalled them with one hand. “Hold on for a second. Let me give you a simple example. Let’s imagine a person named Abe who, whenever he receives messages mentioning kids, responds negatively.” David spread his arms, getting into the story. “ELOPe has to predict how Abe will react to a new email. If the message mentions kids, which Abe historically has a negative reaction to, then it’s a good bet the new email will be received unfavorably. If Abe was your boss, and you were going to ask him for vacation time, spending time with your kids isn’t a good justification.”
He heard a few laughs.
“So there’s no semantic analysis?” Rebecca asked. “We don’t know why he dislikes kids?”
“Correct. We have no idea why Abe feels the way he does,” David said. “We just observe the pattern of behavior.”
“What if my manager hadn’t received any emails about kids?” Sean asked. “How could we predict how he would respond?”
David grinned. Sean knew the answer and was just helping him along. “Let’s say we have another user, Bob, who hasn’t received any messages about children. However, ELOPe groups together Bob, Abe, and about a hundred other individuals based on their almost identical responses to most topics, such as the activities they do on the weekend, the vacations they take, how they choose to spend their time. Let’s say this group of people are ninety-five percent similar. That is, across all the topics they’ve responded to, they are ninety-five percent likely to have the same sentiment in their response: negative or positive. This is what we call a user cluster.”
Heads nodded. David went on.
“If other members of the cluster received emails about kids, and they responded negatively, then ELOPe will be ninety-five percent certain Bob will behave the same way. Of course, situations are rarely so cut and dry, and it is a statistical prediction, which means that five percent of the time ELOPe will be wrong—but most of the time the analysis will be right. So if your boss was Bob, you still shouldn’t mention kids when you ask for vacation.
“Joking aside, ELOPe is working, and we’ve tested the software with users. On average, favorable sentiment in reply emails increases twenty-three percent with ELOPe turned on compared to the baseline. That’s twenty-three percent more vacations granted, twenty-three percent more people agreeing to go on dates, twenty-three percent more people getting their work requests granted.”
Rebecca stared at him. “Wait a second. Going on dates? So you’ve got someone out with another person, someone they wouldn’t have otherwise been with. That seems manipulative, even risky.”
David’s stomach threatened to leap into his throat as his internal danger meter flared into the red. He noticed Kenneth, startled by Rebecca’s objection, leaning over to speak quietly to Sean.
The dating example was so damn controversial. The next few minutes would make or break his project. If Kenneth and Rebecca decided against him, he’d lose Sean’s support and ELOPe would never be released.
“Hold on. Maybe I chose a bad example.” David held up both hands. “Who’s taken a Myers-Briggs personality workshop?”
As expected, everyone raised a hand or nodded in assent. Myers-Briggs or something similar was standard fare for every manager in big companies. “Now, what was the purpose of the workshop? It’s not just to find out whether you’re an introvert or extrovert, right?”
“To work effectively with others,” Sean said.
“Working effectively means what?” David paused. “Learning how others communicate and think, like who is likely to appreciate a data-driven argument versus an emotional argument, or who likes to think out loud versus having time to respond to written arguments.”
He scanned his audience, forcing himself to stay upbeat and chipper even though he feared group opinion could go against the project at this point. “Is that manipulative? Do we take a Myers-Briggs workshop to manipulate people, or do we do it to be able to work more effectively with them and spend less time in arguments and disagreements?”
A few of the VPs turned to Rebecca, waiting for the CEO to respond. She hesitated, then nodded. “It’s helpful. I can see that. I’ve been through more than my fair share of those workshops.”
“And if two individuals took Myers-Briggs together, they’d get to know each other better. Perhaps those people would not only work together better, but as a side effect they might be more likely to have an enjoyable date. What we’re doing with ELOPe is giving everyone the same benefits they would get from an expensive workshop. We’re empowering human beings to become better communicators and collaborators, something everyone wants.”
The tension in the group dropped noticeably, and the audience was once more dominated by nodding heads.
“Remember, we’re measuring sentiment in these messages,” he went on, pacing back and forth in front of the display again. “It’s not a grudging assent: people are having and maintaining more effective and cooperative ongoing communication when our tool is enabled. Once, spellchecking was the b
ig innovation that leveled the playing field between people of good or bad spelling ability. Now we’re leveling the playing field for writing—enabling people of all abilities to create powerful, well crafted communications.”
There was quiet for a minute, then one of the executives asked, “What’s the timetable for releasing this?”
Discussion went on for another fifteen minutes, but the topics were all implementation details and business return on investment questions.
At the end of David’s presentation, Sean walked him to the doorway while the executives helped themselves to another round of coffee and food. “Good job,” he said privately to David, as he ushered him out. “I’m confident they’ll vote to go live when you’re ready.”
As the door closed behind him, David leaned against the wall outside the conference room. The experience had been draining. Then he chuckled. The dating example had been contentious, but it was better to raise the issue and address it early than leave the topic lingering. He was sure the presentation had won them over. The language analysis on his slides that he ran last night in ELOPe predicted a ninety-three percent favorable response.
“Gary, it doesn’t make sense to optimize until after we’re done.”
While David handled the all-important presentation with the bigwigs, Mike patiently defended their resource utilization with Gary Mitchell. Mike wondered, not for the first time, if David had arranged the meeting with Gary to conflict with the executive meeting so Gary wouldn’t attend the briefing.
Mike sighed. Give him a team of developers to motivate, a thorny bug to fix, or a new architecture to design, and he’d be happy. But he hated organizational politics. David owed him one for this.
“Of course,” he went on, “we’ll only use a fraction of the number of servers after we optimize. We’ll work on efficiency improvements when the algorithm is done. Optimizing now would hurt our ability to improve the effectiveness. This is basic computer science.”
“Mike. Mike.”
Mike rolled his eyes at Gary’s condescending tone, a safe maneuver since Gary appeared to be studying the ceiling. Gary leaned back in his chair, arms behind his head, white dress shirt stretched over his belly, jowls hanging down under his chin. Mike wondered how Gary had ended up at Avogadro. He needed only a cigar and ashtray to be at home as a 1950s General Motors vice president.
“I know your project got special approval from Sean to use production,” Gary said. “Those servers are responsible for running Avogadro’s day-to-day operations.” He straightened and stared at Mike, jabbing a fat finger in his direction. “You’re eating up so much memory and bandwidth on AvoMail that I’ve had to twice bring in additional capacity. You know what happens if AvoMail goes down? Millions of customers abandon us, and I get chewed out by Rebecca Smith.” He stood and walked over to the operation dashboard updating in real-time on the wall display. “Hell, I can measure the loss if we slow down by even half a second. You spike CPU usage, we lose revenue.”
“Gary, we—”
Gary tapped the dashboard and ran right over him. “Like every other R&D team, you think your project is manna from Heaven. Meanwhile, I gotta keep things running here, and ELOPe is making us run critically short of capacity. Approval from Sean or not, I’m in charge of Communication Products, and I have ultimate responsibility for ensuring absolutely zero downtime. You’ve got two weeks to get your server utilization down, or I’m cutting off your access to my production servers. And if you have another spike, I’ll shut you down instantly.”
“Listen Gary, we can--” Mike started.
“I don’t want to hear excuses!” Gary shouted. “We’re done. I’ve had this discussion with David repeatedly. Two weeks. You go tell David. Goodbye.” Gary shooed him out of the office with his hand like an errant cat.
Mike stood, then stormed out, blowing past Gary’s startled admin. Resisting the urge to slam doors, he stalked down five floors, fuming with unspent anger. He crossed the street and went down a block, then up again into his own building.
He slowly relaxed during the walk, one benefit of the sprawling campus. Avogadro Corp had expanded so much they now spanned seven city blocks in the northwest part of Portland, on the site of an old trucking company.
As the company and their profits grew over the last fifteen years, they put up one new structure after another, so fast the employees couldn’t keep track of who or what was where. Mike had seen three new buildings go up in the few years since he started.
There was an ongoing curiosity among the employees to discover what the different buildings contained. Most of the campus was quite normal, but there were some oddball aspects, like the telescope observatory opened by randomly chosen employee access cards, and the billiard room that changed floors and buildings. Mike had seen that one himself. Whether the trick was managed by moving an actual room, Facilities staff carting off the contents, or duplication, no one knew. Of course the engineers at Avogadro couldn’t resist a puzzle, so they’d tried everything from hiding Wi-Fi nodes to RF encoding the furniture, with ever more puzzling results.
There was a half-serious belief among some employees that one of the executive team had a Winchester-house complex. Mike had visited the San Jose fixture once, during his college days. Built by Sarah Winchester, widow of the firearms magnate, William Winchester, she had the house under constant construction from 1884 to 1922, believing she would die if the work ever stopped. The idea that a similar belief plagued one of the Avogadro executives, driving them to do the same to the company campus, brought a smile to Mike’s face. On the whole, however, he figured the curiosities were a game to entertain engineers. It takes something extra to retain brilliant but easily bored geeks.
The smile disappeared as Mike crossed the second floor bridge back to R&D and thought of David. He’d blow a gasket at the news of Gary’s ultimatum.
Their recommendation algorithm, which sounded so simple when David explained the idea to a nontechnical person, depended on crunching vast quantities of data. Every email thread had to be analyzed and correlated with millions of other emails. Unlike movie recommendation algorithms, which were clustered using less than a hundred characteristics, it was orders of magnitude more complex to do the analysis on emails. They needed a thousand times more computation time, memory, and all-important database access. Coming out of the meeting, Mike had no doubt Gary had reached the limit as far as available resources were concerned.
Unfortunately, Mike had been less than honest. He shrugged, uncomfortable with himself. When had it become necessary to lie in his job?
The reality was that he, David, and other members of the team had been working for months to optimize ELOPe. Sadly, the current server-consuming behemoth was the best they could do. There weren’t going to be any more efficiency gains, and therefore no way to meet Gary’s ultimatum.
Mike sighed. David would be seriously upset.
A busy morning kept Mike hopping from one urgent issue to the next, despite his desperate need to talk to David. Hours later, with the most critical problems resolved, Mike ran into David’s office before anything else could interrupt him.
“Got a minute?” he asked in a cautious tone, poking his head around the door.
“Of course.”
Like all the engineers’ offices, David’s had room for three or four guests, as long as everyone was friendly and used deodorant. A big whiteboard spanned one wall, and north-facing windows held a view of heavily-wooded Forest Park. Mike was sure the six-month-old setup was less effective for working together than last year’s layout when the team was in one big open space, but he enjoyed the change. Besides, the office would be different again next year.
Mike recounted the meeting with Gary Mitchell. David’s face grew grim before he even finished. “Then he kicked me out. What could I say even if I had the chance? There aren’t any more efficiency gains.”
David sat at his desk, fingers steepled, staring into his darkened monitor. It was a bad sig
n when he stayed in statue-mode for over a minute. Tux the Penguin, the Linux mascot, which Christine had bought David after one of their first dates, wobbled over David’s display in the ventilation system breeze. “So two weeks is our deadline. What do you want to do?” Mike prompted, after he’d endured as much of the painful silence as he could stand.
“Pare down the number of developers we have working on fixes and algorithm improvements,” David said, having apparently reached some conclusion. “How many people can you dedicate to optimization?”
“I’ll focus full-time,” Mike said, starting to count off on his fingers. “Certainly Melanie,” he added, referring to one of their best software engineers. “Figure two or three other folks. Probably five in all. But, David…” He looked him square in the eyes. “We’re not going to make any improvements. We’ve tried everything.”
“All right, starting with the five of you, get focused on performance full-time,” David said, ignoring Mike’s protest. “After we hit our next release milestone on Thursday, we’ll see where we stand.”
Mike sighed and left the office.
Chapter 2
“How’s the project going?” David asked, popping into Mike’s office on Thursday morning. He perched on the windowsill, glancing outside at Mike’s view of the football-field-sized granite calendar that mapped the history of the universe into a single calendar year. A bunch of new hires were getting their orientation on the month of December. Damn, but Mike had the best view from the floor. He turned back to Mike, hoping he’d have some good news.
“Excellent,” Mike said, finally pulling his eyes from the screen. “Everyone finished their tasks for the iteration, code is checked in, and integration tests are running. We’ll know in a few minutes if everything passed.”
Avogadro Corp Page 2