I wrote this in 2017, in my first year as a data scientist. I had come straight out of a master’s program into a start-up lender, and I was the only data scientist there. The post was about how I got over the first set of problems with no one around to ask.
The original title was “From Knowing a Little to Learning a Lot”, and that still describes it well.
Where I started
How fast you grow as a data scientist depends on a lot of things. Your education, your prior experience in the field and the industry, and whether you join a team of data scientists or work alone all shape the learning curve.
For me it was steep. I knew very little about financial services. What I had was a solid understanding of logistic regression, some economics projects involving time series, and a toolkit of R and Excel.
I could have done more to build my skills before starting. But with what I knew, a willingness to learn and a lot of curiosity, I had enough to begin.
My job was to build and maintain the company’s in-house credit scoring models and to produce ad hoc analysis and reports when asked. From day one I faced several gaps:
- I did not know how to build a credit scorecard properly.
- I did not know many advanced analytical techniques.
- I had no way to check that my work met industry standards.
- I did not know how to monitor a model’s effects once it was live.
Here is how I worked through them.
Doing my own research
When you have almost no background in a problem and no one to ask, the first instinct is to research. My master’s program had a lot of independent research built into it, so this came naturally.
For the scorecard, the most important step was understanding the whole thing first and then breaking it into pieces I could manage. I needed to know why a scorecard is used, how it is used, and how it would help the company’s operations.
My research often turned up one of two things. Some solutions were too complex to implement without enterprise software or advanced programming. Others looked too simple to be convincing.
Trying to reproduce projects I found online taught me a lot. It made me much better at R. Along the way I picked up some Python and learned to write queries in Microsoft SQL Server and MySQL, which made my data and model building faster.
Talking to people in the industry
The harder problem was making sure the models followed good practice for the industry. Two things made that difficult.
First, a scorecard for an alternative business lender is very different from the more common scorecards in the industry, such as those for personal loans. Second, subprime business lending as an industry had mostly grown up after the 2008 financial crisis. Published research was thin, and most of the know-how was proprietary.
So I did more research, but more importantly, I talked to people who worked in the industry and took what I could from those conversations. Most of them were about which techniques were widely used. LinkedIn and my own connections helped a lot here. I learned how to reach out to professionals online and how to connect ideas from different people inside my own company.
Trial and error
There is a lot of pressure when you start, with deadlines and expectations. The scorecard was my first project, and with what I knew, I was pushed into trial and error.
At first I was in a loop of researching and rebuilding, updating the scorecard whenever I learned a better practice. There was very little feedback from internal users because everyone assumed the model was working as it should.
What made this work was constant communication with the rest of the company. I learned a lot about the technical side of model building. I also learned that a standalone data scientist has a particular place inside an operations team.
Being prepared and building confidence
Hard data science problems are not solved quickly. When you work alone and are expected to make decisions on your own, you run into worry and frustration. When something goes wrong with a model, you are the first person accountable, and that can be unsettling.
I came to accept that these feelings were normal. What helped was being prepared. I tried to have thorough answers ready for questions from the company and to address concerns directly. When someone raised a problem with a model or a method, I came back with an answer or a fix.
I also tried to be honest about what I could and could not do. I did not know everything, but I wanted to learn so I could do my best work and help the company grow. With open communication and support from senior management, the personal worries faded, and my learning of applied business data science sped up.
Moving forward
What I valued most about those early days was how much I learned and how much I grew as a person. The learning does not stop. New modelling needs come up, the data keeps growing, and new techniques arrive with new technology.
In a fast-growing company, solving one problem tends to uncover ten more. The best part of being early in my career was knowing I still had a lot to learn, and being happy to take the challenges one at a time.
Looking back from 2026
The four habits in this post still hold: research the problem, talk to people who have solved it, iterate in the open, and be honest about what you know. None of them depend on a particular tool.
The research step looks very different now. In 2017 I spent weeks piecing scorecard methodology together from blog posts and forum threads. Today an LLM can explain weight of evidence, draft the binning code and point you at the standard references in an afternoon. That is a real speed-up. It also makes it easier to ship something you do not understand, so the step of breaking the problem down and checking each piece matters more, not less.
Entry-level work has shifted too. A lot of the tasks a junior data scientist used to learn on, like writing first-draft SQL, cleaning a CSV or producing a routine report, are now partly done by AI tools. New data scientists are expected to review and judge that output earlier in their careers, which takes domain knowledge they may not have had time to build.
Being the only data scientist at a small company is still common, and still hard. What I would add to my 2017 advice is to find a peer group outside the company early, whether that is a meetup, an online community or a former classmate. It is the closest substitute for the team you do not have.
The part about accountability has not changed at all. When a model goes wrong, someone has to explain it. Being prepared and being honest are still the best way to earn the trust to keep building.