This Data Neighbor Podcast episode interviews Karthik, a senior machine learning engineer at Meta who previously worked at Amazon and interned at Google, and argues through his story that a move into machine learning is a sequence of deliberate, adjacent steps rather than a lucky break. He ends on two claims: that his apparent confidence is imposter syndrome turned into a habit of seeking out smarter people, and that skydiving taught him every problem is solvable.
Getting in the door. Karthik's interest began around 2013 or 2014 with Google search and the PageRank algorithm, which led him to data mining and machine learning courses. From Jaipur he cold-emailed alumni and professors; he estimates 99% never answered. He interned with a professor at another Indian university to find contacts, and eventually a Minnesota alum's lab took him on, which led to a master's at the University of Minnesota in 2015. He calls himself an introvert, and credits a mentor's line that the worst case is a "no" that leaves you where you started. The hosts' reading is that what looks like luck was a 1% chance worked for relentlessly.
The detour through software engineering. A student loan, a visa clock of a few months and the dominance of PhDs in ML hiring at the time pushed him toward software engineering. He chose an Amazon pricing team next to an ML team over a better-paid Google Android offer, citing Bezos's regret-minimization framework. Once there he kept asking his manager for ML work, and later declined a fast track to senior software engineer in about four years, taking smaller ML projects instead because a senior-level switch would carry higher expectations.
Building the path. He contributed a feature the scientists lacked the bandwidth to build; it went into the next model iteration and built the team's trust in him. His route ran through ML infrastructure (an inference platform for dynamic pricing), feature engineering, ML engineering, search within Amazon, and then Meta. He says most software engineers who move to ML follow this route because infrastructure skills transfer. For someone outside tech, such as in healthcare, he stresses domain knowledge as the hardest part to earn, then coding and a network of practitioners; the hosts frame it as a series of bridges, and Karthik adds that some people have to switch companies to make a jump.
What the job is. He describes ML engineering as a spectrum from infrastructure to research, with titles that differ by company and team. As an applied engineer he owns search and search ads: diagnosing why some queries fail, then choosing between a heuristic, an optimization or a model, and evaluating through A/B tests.
Confidence and perspective. Karthik says he struggles with imposter syndrome at every new team, and would again in a field such as computer vision, and that it pushes him to find people smarter than him. He took up skydiving in 2020 and is now licensed. He says seeing the world from 13,000 ft eased his anxiety, and he recommends a tandem jump to everyone, putting the risk of death at under 0.001%.
In this Data Neighbor Podcast episode, two of the three co-hosts (the channel owner is absent) argue that impact is any work that changed something useful to the company — a metric, a roadmap or a process — and that data people lose credit for it mainly by never following up. They close on a recipe for any data role: track it, follow up, and state it concisely.
What impact is not. The co-hosts open with counter-examples: a cutting-edge tool such as XGBoost or an LLM used on a problem with no business return, and an ad hoc request answered and handed off with a "thank you", after which nothing is heard. The second may have had impact, but impact you do not know about cannot go on a resume or into a recruiter screen.
Three kinds of impact. One co-host's framework: moving a metric through features built on your insight, changing a roadmap (an analysis that shuts down a hopeless business line), or changing a process (a one-click experiment-analysis tool). The other co-host maps these to revenue gained, money saved, and productivity, which is the most indirect link.
Mindset. The co-host who manages a data-science team says they learned about impact only when recruiters asked for numbers. Speaking about it gets you more of that work, and an engineering director's puzzle analogy holds that the insight is the piece without which there is no impact. Credit is shared, weighted among contributors, not split hair by hair.
Tactics by role. Follow up with PMs and engineers on experiment ramps and results, document them, and book a 15-minute check-in the next week. Analysts describe dashboards by scope and audience, as in their example line: a company-level dashboard used daily by executives with over 500 views. Product data scientists cite the roadmap and metric movement, ML engineers the model and the metric, and experimentation or research teams the company-wide process change.
The first episode of the Data Neighbor Podcast, hosted by three former Nextdoor colleagues, argues that there is no single route into data work: a break snowballs for whoever has prepared for it, and the field's titles are messier than its few underlying families.
Why the podcast. The hosts frame it as demystifying data science, a job that means something different at every company, for people without friends in the industry. They promise weekly episodes, guests from ICs to VPs, and a coming episode on why there are so many data roles.
Three routes in. One co-host went from computer science and SQL-heavy software work to a business-school master's, six years at Microsoft, then eBay and Nextdoor, and is moving to Grammarly as a senior manager. Shane went from policy and the Peace Corps through PwC and AppFolio to senior data scientist at Stripe. The third studied statistics, skipped the actuary path, and went through gaming to Facebook, Pinterest, LinkedIn and director at Nextdoor.
Luck and openness. The hosts argue that one break is all it takes and that you raise your "luck surface area" by working on yourself, citing a top-10% resume list and certifications behind a Microsoft offer. Shane adds that people should try different data roles and company sizes, because the skills transfer.
Titles. A host says the field's hundreds of titles fall into five or six families and need a cleanup. Examples include a Microsoft title of "applied statistics engineer", and an argument on LinkedIn over whether causal inference counts as data science.
In this Data Neighbor Podcast episode the guest, a senior director of data science at Instacart who previously led a data-science team at Facebook, argues that breaking in takes preparing for luck and then exploiting it hard. He lays out two of the title's five steps (skip prestige targets, and invest about two hours in each application) and ends on the view that AI will make data scientists more productive rather than replace them.
His path. He studied mechanical engineering, found it dull, and took a master's in engineering management at Duke. That led to a business analyst job at Capital One in 2012, doing modeling and experiments before the data-scientist title existed there. Bored by credit cards, he applied widely in tech and Google called back six months later; the move was technically a demotion from L4 to L3, which he did not notice at the time. In 2016 Facebook turned its product analysts into data scientists, and he joined as an L4. He stayed almost six years in one seat, growing it into a team of about 25 with managers of managers, before leaving for Instacart about two and a half years before the recording.
Luck. He splits luck into a before and an after: put yourself in position, then go flat out when it lands. His highway analogy says that when the traffic clears you floor it, and he holds that someone whose area becomes the CEO's focus should work 12-hour days. Interviews are random too, he says, and a defeated or entitled attitude shows through to recruiters in a market he calls the worst he has seen for new grads.
The two steps. First, avoid big-name companies and titles without a connection, and look locally for analytical roles. Second, judge success by offers, not applications: spend about two hours each, tailoring the resume and finding connections, because his junior postings draw hundreds of LinkedIn messages and low-thousands of applications. A host adds that applicants should ask what they can give. The guest says that could be the good feeling of helping, a specific question about the person's work, public bug-finding, or open-source work. A host who once analyzed LinkedIn's job-seeker data recalls referrals being roughly 50x as effective as blind applying. The guest's tip is to interview as though talking to a coworker.
Discovering a strength. At Facebook he was asked to lead a monthly business review with the head of News Feed when the usual presenter was out, found he was good at telling data stories, and built it into a staple of how the business was run. That gave him direct time with leaders. His rules: do the chewing for executives, and make a deck clear enough that a seven-year-old could explain it to another, because executives retell it. A host adds the rule of cutting any artifact in half twice.
AI. The guest calls the claim that AI will replace data scientists untrue, likening AI to unlimited intelligent college interns. He says Instacart's data scientists each have 1.5 to 1.8 people's worth of work, so help from AI gets more done rather than eliminating jobs. He uses it for writing — shortening and polishing in his own style after feeding it samples of his writing — and for learning unfamiliar domains such as marketing, but finds its behavioral-interview answers canned. A co-host warns that AI-written Slack messages are obvious.