When you click on a sponsoring school or program advertised on our site, or fill out a form to request information from a sponsoring school, we may earn a commission. View our advertising disclosure for more details.
Gautam Sihag is a senior machine learning engineer at Shipt, where he leads the development of machine learning systems that determine accurate pay for delivery contractors and protect the company from financial losses across its contractor network. It is high-stakes, production-grade work that requires ML models to perform reliably at scale, day in and day out.
Sihag has been working at the intersection of data and engineering for over a decade. He began his career as a data engineer at Altice USA before joining Target as a senior AI engineer, where he built and maintained demand forecasting models spanning thousands of store locations nationwide. He joined Shipt in 2021, where his focus has expanded to include agentic AI development and the foundational infrastructure that allows engineers across an organization to interact with large language models safely and effectively.
He holds a master of science in computer science from Columbia University, where he studied machine learning, big data analytics, and cloud computing. He also served as a course assistant for Columbia’s big data analytics course and conducted research for the university’s History Lab.
AI can feel like a vast, shapeless force, something that is happening everywhere and nowhere at once. But for engineers trying to understand their place in it, that ambiguity is not especially useful. Sihag breaks the landscape down into four distinct verticals, and knowing which one your organization operates in determines what AI actually looks like for you on a daily basis.
The first vertical is the one that gets the most attention: the organizations actually building the underlying models. “They are organizations like Google and OpenAI. They have already built the underlying model,” Sihag explains. “A lot of organizations wanted to build their own LLM, but now it has been commoditized, and everyone is using it as a service rather than building something in-house.” That race, for most companies, is already over.
The second vertical is less visible but increasingly critical. These are the teams building the internal platforms that allow everyone else in an organization to interact with large language models safely. Without that layer, engineers and analysts are left to interact directly with external tools, often without any controls around what proprietary information gets shared. “We don’t want anyone and everyone to just directly go and start using LLM,” Sihag says. “We have to build foundation layers so that it enables everyone else to use or interact with it safely.” He compares it to the situation hospitals face when considering AI tools: the capability is there, but safeguards for sensitive data must be built before any of it can be used responsibly.
The third vertical involves building agentic applications on top of those foundation layers, software tools that can reason, plan, and take action across multiple steps without constant human input. The fourth is perhaps the most widespread: everyday engineers and knowledge workers simply using AI tools to get their work done faster. Sihag himself operates across both. He estimates that roughly 20 percent of his time currently goes toward using AI tools to assist his own work, while another 25 to 30 percent has been dedicated over the past six months to building the foundational capabilities that will allow others in his organization to do the same. “It’s still very new,” he says. “A lot of organizations, except the major players, are still on their journey to try to understand how it can be used to create value against the cost.”
The most tangible change Sihag has seen in his day-to-day work is speed. AI agents can now write end-to-end code in a development environment, and what used to take a team of engineers two weeks can realistically be completed quickly. “What we used to build in two weeks can now be shipped effectively in about two to four hours if someone knows how to correctly do it,” he says. He has friends at Pinterest who describe even more dramatic compression: work that once required two to three engineers over two weeks now takes only hours.
But speed is only part of the story. The nature of the work itself has shifted. Engineers are spending less time writing code line by line and more time on architecture and direction-setting, crafting detailed instructions that tell the model exactly what to build and in what order. “It’s not like you just write a line of what you want and it will build everything,” Sihag says. “That’s an incorrect explanation.” The prompts that actually produce good results are long, precise, and carefully structured, more like a project brief than a search query.
There is also an unexpected dynamic emerging around learning. When a model is set up correctly, it pushes back. It tells an engineer when their line of reasoning is off, or when there is a better approach. “That student-teacher mentality has sprung up,” Sihag says. “We are corrected more often than before.” For engineers who have spent years building expertise in a particular way of doing things, that can be a disorienting experience, and a genuinely useful one.
The speed gains are real, but they come with a catch that Sihag says doesn’t get nearly enough attention. When code is being generated at this pace, quality control becomes a serious problem. “There’s a lot of slop building up,” he says. “A lot of unwanted code or unwanted features have been shipped.” In traditional software development, nothing reaches production without peer review. That standard hasn’t changed, but the volume of code needing review has grown enormously. “I’m happy to review 500 lines of what is being done,” Sihag says. “But if I have to review thousands of lines of code every day, that’s a problem.” One wrong feature, he notes, can cost a company millions.
Hallucination is the other persistent challenge. LLMs are prone to filling in knowledge gaps with plausible-sounding but incorrect information, and managing that tendency requires constant attention. Sihag describes it in terms of a context window: up to about 40 percent of its capacity, a model performs well. Beyond that threshold, reliability drops sharply. “It just goes haywire and will say anything it likes,” he says. Preventing that requires engineers to build supporting tools that provide the model with the right context at the right time, and to explicitly instruct it when to stop and seek additional information rather than improvise.
All of this points to what Sihag identifies as the hottest topic in the field right now: context engineering. To understand why it matters, it helps to know how LLMs work. Every conversation a model has exists within a context window, essentially the total amount of information the model can hold in memory at once and reason from. That window fills up faster than most people realize. Every previous output, every database query, every tool the model calls adds to it. Once it is full, performance degrades.
Context engineering is the discipline of carefully managing what goes into that window and in what order, so the model always has the most relevant information available and none of the noise. Getting it right is not simply a matter of writing better prompts. It requires a deliberate, architectural approach to how information flows in and out of the model at every stage of a task.
The excitement around AI in engineering is genuine, but so is the exhaustion. Sihag is candid about the toll the pace of change is taking on the people living through it. A year ago, keeping up with the field meant steady, manageable upskilling. Now it means something closer to a second job. “If I’m not working, I’m equipping myself so that I know what’s going on in the market,” he says. “We’ve lost our relaxation time. The time when we could enjoy that we know something, and we can contribute or build a career around it, that has gone.”
The emotional relationship engineers have with their work has shifted, too. Software engineering has always attracted people who are genuinely passionate about building things, who derive satisfaction from watching a complex system come together over days or weeks of focused effort. That slower, more invested process is largely gone. Work that once took two weeks of careful, iterative development now takes hours to generate, and the engineer’s role becomes one of reviewer and corrector rather than builder. “The burnout is real,” Sihag says simply.
And yet, he and many of his peers keep going, largely because the moment is genuinely thrilling. “The last eight months have been really exciting,” he says. “We have not seen things come up so fast, at least in the eight to ten years I’ve been working in the industry.” For now, most engineers are choosing to lean into that excitement rather than reckon with what it is costing them. But Sihag is honest that it is not a sustainable equation and that, at some point, something will have to give.
For students currently working toward a computer science or engineering degree, Sihag’s advice is grounded and a little counterintuitive, given all the hype around AI. The fundamentals still matter. Deeply. “Agentic engineering is still software engineering,” he says. “It’s just faster.” Understanding how operating systems work, how network layers function, and how data structures are built is not a skill that AI has made obsolete. They are the skills that allow an engineer to evaluate what AI produces and catch what it gets wrong. “We still need to evaluate it,” he says. “At least for the next two years, knowing data structures is still crucial.”
What has changed is the range of skills engineers need beyond the technical ones. Sihag points to product thinking as the most significant gap he sees in the field’s current training of its next generation. Engineers who understand the business context around what they are building, who can think about why a feature matters and what problem it actually solves, will have a meaningful advantage over those who can only think in code. “A computer science major should start knowing a little more product,” he says. “Having some business knowledge and trying to understand how and why the world works in a certain way is a skill that will become crucial.”
He also acknowledges that the industry itself needs to adapt. The way engineering candidates are currently evaluated, through intensive coding challenges and algorithmic problem-solving exercises, was designed for a different era. As AI handles more of that work, those benchmarks will need to change. “That has to come from the industry,” he says. It is, he admits, a missing piece that no one has quite figured out yet.
If there is one thing Sihag is certain of, it is that the landscape will look different in another six months. New protocols, new tools, and new terminology are emerging in a cycle that makes it nearly impossible to build stable expertise around any single approach.
Model Context Protocol, or MCP, was the dominant framework for AI agents’ interaction with external tools just a few months ago. The conversation has already shifted to a new concept: “Every two weeks we have to look at what is new and how it would help,” Sihag says. “Because we had built something very fast over the last three months, it also means there’s a lot of slop building up.”
For organizations still trying to find their footing with AI, that pace is both the challenge and the point. The companies that are getting the most out of these tools are not necessarily the ones that moved first. They are the ones who thoughtfully built, invested in foundational infrastructure to use AI safely, and resisted the urge to ship everything as fast as the tools would allow. The cost of moving too fast, in code quality, in security, in engineer wellbeing, is real and it is accumulating.
Sihag is optimistic, but his optimism is measured. The tools are genuinely powerful, the productivity gains are real, and the pace of innovation is unlike anything he has seen in his career. But he is also watching closely as the tradeoffs pile up, and he is not convinced the industry has fully reckoned with them yet. “We are trying to understand how AI can be used to create value against the cost,” he says. For most organizations, that question is still very much open.
With a keen understanding of the challenges and opportunities in higher education, Kimmy Gustafson regularly contributes insightful articles to OnlineEngineeringPrograms.com, providing readers with a comprehensive view on the evolving landscape of engineering education since 2019. She has interviewed many engineering experts on a range of subjects, including geoengineering.
Kimmy has been a freelance writer for more than a decade, writing hundreds of articles on a wide variety of topics such as startups, nonprofits, healthcare, kiteboarding, the outdoors, and higher education. She is passionate about seeing the world and has traveled to over 27 countries. She holds a bachelor’s degree in journalism from the University of Oregon. When not working, she can be found outdoors, parenting, kiteboarding, or cooking.
Today, digital twins are not limited to just physical objects. With the rise of virtual and augmented reality technologies, digital twins can now replicate entire environments and systems in a virtual space. This has opened up new possibilities for testing and simulation, allowing companies to reduce costs and risks associated with physical prototypes.
Diversity and inclusivity aren’t purely idealistic goals. A growing body of research shows that greater diversity, particularly within executive teams, is closely correlated with greater profitability. Today’s businesses are highly incentivized to identify a diverse pool of top talent, but they’ve still struggled to achieve it. Recent advances in AI could help.
The ability of a computer to learn and problem solve (i.e., machine learning) is what makes AI different from any other major technological advances we’ve seen in the last century. More than simply assisting people with tasks, AI allows the technology to take the reins and improve processes without any help from humans.
When is a computer more than a machine? One person’s concept of artificial intelligence (AI) can be quite different from the next person’s. In its most basic form, artificial intelligence is a process in which machines take cues from their environment and act responsively to achieve a goal.
Not long ago, self-driving cars were science fiction. Today, not so much. Influential companies like Tesla, Uber, Apple, and Google boast dynamic auto-drive programs, and many new startups are following their lead.
