We have a need to hire more AI engineers at Gainwell – so I want to put together my brief notes on what types of skills I expect those individuals to have. Many grad programs in data science have not caught up – the work involves very little fitting of models (either tree-based or deep learning) and is more about integrating LLMs into various products.
Understanding the LLM APIs
While people often denigrate AI engineers as just gluing API calls together, the LLM APIs are some of the most complicated APIs folks will ever end up interacting with. For just a basic rundown, the things you should understand are:
- how traditional chat conversations work, e.g., system prompts, user input, and agent responses
- how tool calling works (and the difference between server-side tool calling and client-side tool calling with locally defined functions)
- how caching works and how to estimate costs with LLM APIs
There are variations of these across different frontier model providers, but they are pretty similar in nature.
Structured Outputs and Document Processing
One of the most common tasks for using AI is replacing laborious document processing – imagine getting a PDF in an email, and the AI automatically extracts the information from the PDF into a spreadsheet for you.
There are still many people who do that manually! (And in healthcare we still often use faxes.) This involves understanding structured output extraction.
While it may seem simple to identify a schema for what you want and just have the LLM go brr, I still need software engineers to help put together the pipes. Regular questions for this will be:
- Should we do extraction from images/scans, or from OCR’d text?
- Can we process an entire document at once, or should it be sliced up first?
- If we use images, what is the resolution?
- Should this be batch or real-time? Are the TPM limits for our account sufficient?
- What is the cost/accuracy for different LLM providers?
This does actually involve one skill many data scientists will have experience with – evaluating the accuracy of different tools, false positives vs false negatives, etc.
How to accomplish agentic tasks
Part of being an AI engineer is helping others get the most out of agentic workflows. I know my devs know how to set up Claude Code to do various tasks. We need to help other neophytes (often not coders) to set up agentic systems to help them out with various tasks.
Sometimes this will just be “here is how you need to set up your documents, and then just ask Codex to review X and do Y”. Sometimes it will be “you are spending too much on Astra, you should be using Luna for this”. And sometimes it will be “we should build some more robust software and a user interface to help people with this specific task”.
Imagine a nurse having to do a complex audit of a 500-page medical record to see if the medical bill was correct. We need to figure out how to use AI to help the nurse do their job better and faster.
This may involve a pre-review pipeline (answering standard questions for the nurse to review). It may involve doing named entity recognition and OCR on the document, so it is easier to search for specific materials. Should we build a chat interface? Do we need to redact PII before it goes to the AI tool?
Another common application at my work is using agentic AI tools to conduct automated data analysis. Imagine you have an automated alert for a log if there is a spike in latency. You can have AI automatically go and do subsequent analysis to try to understand the root cause of the issue.
This requires understanding how the APIs work, setting up MCP servers or tools in a sandboxed environment, and setting up permanent memory traces for subsequent analysis. (In addition to knowledge of outlier detection methods.)
It is particularly tricky to build good benchmarks for these. How do you know the system is working? For the most part, these applications need to be clearly more accurate than humans or people will for the most part just dismiss them.
You still need to know Python
Finally, while I know that agentic tools can write basically 100% of the code, I still need you to know the basics of Python. I need you to know when you should use a function, or make a general class. I need you to know when it makes sense to run a particular task in parallel. I need you to know the difference between a string and a datetime object.
You should be using agentic AI as a tool, not as a crutch, to help you do your job. Not knowing the basics will make you obsolete relative to current standards of AI coding agents. If you know Python well, though, the agent coding tools can essentially be force multipliers to help you spin up your ideas much faster.
How to learn?
I specifically wrote LLMs for Mortals to address this gap in current education. You can pick up an EPUB or print copy (use promo LLMDEVS for $25 off the EPUB).

I also have an entry-level Python book, Data Science for Crime Analysis with Python. (You can use promo code TWOFOR1 to get $30 off your total order when purchasing two books.) While going for a degree is good, you can ultimately learn these things on your own (and I am happy to hire good self-taught programmers at Gainwell for these positions).
Beyond just book learning though, pick up a project to add to your portfolio. Learning by doing is one of the best ways to learn as well. And the LLM costs for processing are really cheap, so you can put together, say, a nice document processing pipeline demo for well under $1 in API costs. Very few individuals who apply for jobs have a good portfolio, so showing some independent work (that is not just a class assignment) goes a long way for me.
For a good exemplar of potential AI pipeline demos, check out Gio Circo’s blog. He has plenty of recent LLM pipeline examples that are pretty relevant for the work we are doing at Gainwell.
