Career Change to Data Scientist at 40

Why did you leave teaching?

I loved teaching. I need to say that first because people hear "teacher who left" and they assume burnout or low pay or both. The pay was a factor, yes. I was making $74,000 after fifteen years with a master's in education. My husband Derek is a physical therapist, he makes about the same. We have two kids. We were fine but we were never going to be more than fine. And I wanted to be more than fine.

But the real catalyst was a summer program I did at Rutgers in 2023. They offered a six-week data science bootcamp for STEM teachers, funded by an NSF grant. The idea was that teachers would bring data literacy back to their classrooms. I went because it was free and it was summer. By week three I was staying up until midnight doing the homework. Not because it was hard. Because it was interesting in a way that grading lab reports hadn't been for a long time. I hadn't felt that kind of intellectual engagement since I was in grad school for chemistry. The woman running the program, Dr. Weisman, she pulled me aside during week five and said "you should consider doing this professionally." I laughed. I was forty years old with a teaching certificate. But the idea stuck.

What did the transition look like?

Long. I taught for one more year while doing Georgia Tech's online master's in analytics, which is designed for working professionals. Classes at night, assignments on weekends. Derek took over bedtime routines. My daughter Ariel, she was 9 at the time, started telling people "mom does homework too." The master's took 22 months. During that time I also did three Kaggle competitions and built a portfolio project using publicly available clinical trial data from ClinicalTrials.gov. That project is what got me my current job. The hiring manager, a woman named Sutton, said she'd seen fifty applications with bootcamp projects using the Titanic dataset. Mine was the only one that used pharmaceutical trial enrollment data to predict site activation timelines. She asked me how I chose that dataset. I told her I'd been reading about clinical trial inefficiency and wanted to see if enrollment patterns were predictable. She hired me.

What transferred from teaching?

Three things, and they're not the ones people guess. The first is experimental design. I taught AP Chemistry for fifteen years. The entire course is structured around designing experiments, controlling variables, interpreting results, and understanding sources of error. That's statistical thinking. When I build a model, I think about it the same way I thought about a lab experiment: what are you testing, what are you holding constant, what could confound your results? My manager Sutton told me once that I design experiments better than people with statistics PhDs. I think that's generous, but I also think she's partly right, because in a classroom you learn to design experiments that a seventeen-year-old can execute correctly. That constraint makes you very precise.

The second thing is explaining things to non-experts. In teaching, your entire job is taking something you understand deeply and making it accessible to someone who doesn't understand it yet, without losing the important nuance. That's exactly what happens when I present model results to clinical operations managers who aren't data scientists. Last month I explained a survival analysis to a site director and he said "oh, so it's like a weather forecast for enrollment?" And I said "yes, exactly like that." In my head I was thinking about censoring and hazard functions, but the weather forecast metaphor was what he needed to make a decision. A teacher's instinct is to find the right metaphor. That skill is worth more in business than any Python library I know.

The third thing is patience with process. Teaching is the most bureaucratic profession I've ever experienced, and I say that now having worked in pharmaceutical analytics, which is also extremely bureaucratic. But teaching taught me that sometimes you just fill out the form and move on. My colleagues who came from tech startups get frustrated by the validation processes and documentation requirements in pharma. I don't. I spent fifteen years doing lesson plans, IEP meetings, standardized test prep schedules, parent-teacher conferences. Filling out a model validation document is nothing.

What had to be unlearned?

The biggest thing was my relationship with certainty. In teaching, you need to be certain. When a student asks "is this the right answer," you can't say "it depends on how you define right." You say yes or no. You grade with a rubric. There's a key. In data science, almost nothing is certain. My first few months, I kept wanting to give definitive answers. "The enrollment rate will be 4.2 patients per site per month." Sutton would say "what's the confidence interval?" And I'd give it but I'd feel uncomfortable because the interval was wide. She told me "the interval IS the answer." That took months to internalize. In a classroom, uncertainty is a sign that you didn't prepare well enough. In data science, uncertainty is the honest output. Communicating uncertainty clearly, without hedging so much that the business can't act on it, is a skill I'm still developing.

What about the salary change?

I went from $74,000 as a teacher to $105,000 in my first data science role. After two years I'm at $118,000. The difference is life-changing. Not luxury life-changing. We're not buying a vacation house. But "we can fix the roof without dipping into savings" life-changing. "I can contribute to the kids' college funds" life-changing. Derek and I went out to dinner last month and I didn't check the prices on the menu. That's the first time in maybe seven years. The thing people don't tell you about teacher pay is that the limitation isn't the number itself. It's the ceiling. After fifteen years at $74,000, I could see exactly where I'd be at twenty years: $82,000. There was no path to meaningfully more. In data science, the ceiling is significantly higher and it's not capped by a union scale. That optionality is what I was really looking for.

The part nobody talks about

I miss the noise. That's the thing I didn't expect. A classroom has energy. Thirty teenagers have opinions and questions and they argue with each other and they laugh and sometimes they surprise you with something brilliant you didn't anticipate. My office now is quiet. I sit in a room with my laptop and I write SQL and I attend meetings where people are polite and measured and nobody raises their voice or says "but WHY does the electron do that?" I have a good job and I'm grateful for it and sometimes at 2 PM on a Tuesday I miss the chaos of fifth period. Not enough to go back. But enough to notice.

What made you leave operations?

I didn't leave operations. Operations left me. The plant I managed for six years got consolidated into a facility in Tennessee in 2023. They offered me a relocation package. Jolene and I looked at each other and both said no. She's a dental hygienist with her own patients. The kids, two boys, were in middle school. We weren't moving to Chattanooga.

I had a severance package that covered about eight months. And I'd been doing something that my HR department would have called "unauthorized data science" for about three years. I'd built Excel models to predict weekly production output based on raw material delivery schedules, machine downtime patterns, and seasonal order fluctuations. I'd built a system in Google Sheets that tracked defect rates by shift, by machine, by operator. When I told our plant engineer Rudy about the defect rate tracker, he said "you've basically built a statistical process control system in a spreadsheet." I didn't know what statistical process control was at the time. I Googled it and spent the rest of the day reading about control charts. That was probably the moment I realized I'd been doing a version of data science without knowing the vocabulary.

How did you make the transition?

Bootcamp. I did a 14-week data science bootcamp in Dallas. Galvanize, which is now owned by Hack Reactor. I was the oldest person in my cohort by about twelve years. Most of the other students were 26 to 30 with bachelor's degrees in unrelated fields. A few had computer science backgrounds. I had a mechanical engineering degree from UT Arlington that I'd gotten in 2004 and hadn't thought about in a decade.

The bootcamp was... humbling. The first two weeks were Python fundamentals and I was fine. Loops, conditionals, functions, data structures. Engineering school gave me enough programming background to keep up. Week three was pandas and NumPy and I started falling behind. Not because the concepts were hard, but because everyone else typed faster and knew keyboard shortcuts I didn't and could debug a traceback in thirty seconds while I was still reading the error message word by word. I stayed after class every day. I did the exercises twice. By week six I was keeping pace. By week ten I was helping other students with the statistics modules because it turned out that my years of tracking defect rates and production variance had given me an intuitive understanding of distributions and hypothesis testing that the younger students were seeing for the first time.

What transferred from the factory floor?

Domain knowledge in operations. That's the answer, and it's why I got hired. My current company builds route optimization software for regional delivery fleets. Twenty to fifty trucks, mostly food distribution and building materials. When I interviewed, the CTO, a guy named Mitchell, asked me to explain how I'd approach predicting delivery time windows. I started talking about vehicle capacity constraints, loading dock scheduling, driver hour-of-service regulations, and the difference between weight-limited and volume-limited trucks. He stopped me after about three minutes and said "you're the first candidate who knows what a cube-out is." A cube-out is when a truck is full by volume before it reaches its weight limit. It happens all the time with lightweight, bulky products like insulation or packaged snacks. If your demand forecast doesn't account for cube-outs, your route optimization is wrong because you're assuming you can load more stops than you actually can. I knew that from fourteen years of watching trucks leave a loading dock. The CS grads he'd interviewed before me didn't.

The other thing that transferred was comfort with messy, incomplete data. On the factory floor, you never have perfect information. The sensor on Machine 7 has been miscalibrated for a week. The shift log from Saturday is missing because the operator forgot to fill it in. You learn to make decisions with what you have and account for what you don't. In data science, people get paralyzed by missing data. They want to impute every null value and handle every edge case before they'll run a model. I just run the model and then figure out if the missing data matters. Usually it doesn't. When it does, I fix it and rerun. That iterative, pragmatic approach came directly from manufacturing, where you can't stop the line to wait for perfect data.

What had to be unlearned?

Speed. In operations, speed is everything. You make a decision and you implement it and you see the results by end of shift. If it's wrong, you adjust tomorrow. In data science, at least at a startup, things move fast compared to a factory, but the feedback loop for model performance is weeks or months, not hours. I built a demand forecast in my first month that looked great on historical data and was wrong by 30% in the first live week. I panicked. My manager Kira said "that's a first iteration. Now we figure out why." In the plant, being wrong by 30% on a production estimate would have been a crisis. Here it was Tuesday. That recalibration of what "wrong" means and how long you have to fix it took me a while.

The other thing I had to unlearn was my instinct to fix the process, not the model. In manufacturing, when output is wrong, you fix the machine or the workflow or the training. In data science, when the model is wrong, you usually fix the model: add features, try a different algorithm, retune parameters. My instinct was always "the data pipeline is broken, let me fix the pipeline." Sometimes I was right. But sometimes the pipeline was fine and the model just needed more features or a different architecture. Learning to diagnose which layer has the problem, data, features, model, or evaluation, took about six months of getting it wrong in various combinations.

What's the salary reality?

At the plant I was making $92,000 base with a $6,000 annual bonus. Total comp about $98,000. At the startup I started at $95,000, which was a slight pay cut. After 18 months I'm at $108,000 plus equity that's worth whatever the company decides it's worth, which right now is theoretical. The equity is 0.15% of the company. Mitchell told me that if they hit their Series B targets, that could be worth $60,000 to $80,000 in a couple years. Or it could be worth nothing. I don't count it until it's liquid. Jolene and I agreed on that.

The financial math of the career change was: eight months of severance, $15,000 for the bootcamp, six months of job searching (during which I burned through most of the severance), and then a starting salary roughly equal to what I'd been making. So the transition cost us about $15,000 in bootcamp tuition and maybe $20,000 in lost income during the search period. We'd saved enough to cover it. The return on that investment will take about three more years to materialize in salary growth, assuming I stay on the data science track. But the option value is significantly higher than what I had in operations management, where the ceiling at my level was about $110,000 without becoming a plant director, which requires relocation and 60-hour weeks.

The part nobody talks about

Being new again. I was 42 years old and I was asking a 26-year-old named Sage to explain how to set up a virtual environment in Python. Not because I couldn't Google it, but because she could answer in thirty seconds and she was sitting right there. Sage was patient and helpful and never condescending. But there's a specific feeling of being senior in age and junior in knowledge that nobody prepares you for. At the plant, I was the person people came to with questions. Rudy, the shift leads, the quality inspectors, they all looked to me for answers. Now I'm the person asking questions. And the answers come from people who are fifteen years younger than me and have been doing this since college. I'm not ashamed of asking. I'm just... aware of the inversion. It gets a little smaller every month. But it hasn't disappeared.