Ro.co: there are basically no interview reports for this company on the forum, so I'm sharing a data point. It's based in NYC, a telehealth startup that's been around for several years, and its main competitor is Hims & Hers.
I interviewed for a Senior Applied AI Engineer position. My background is traditional SWE, and I applied blindly online. I talked to the recruiter and the HM separately first, then did a 1-hour take-home assignment on a platform called Aptora. It was my first time seeing it, and I couldn't find anything about it online. The platform has a built-in AI Assistant where you can ask questions and write code.
The problem wasn't hard. There's a service that classifies patient information. For example, if a patient says "my throat hurts", it gets classified as "upper respiratory", and then the patient is routed to the matching department/doctor. The team says the classifier isn't very accurate and we're asked to debug it. Three API endpoints were provided:
triage(msg: str) -> dict
get_categories() -> list[str]
health_check()
They also provided an LLM API, to be called if necessary. The code for all three endpoints is very simple. The core is triage, and the dict it returns contains message, category, urgency, and confidence_level. The problem also came with a Markdown file that asks you to describe what bugs you found, your prioritization, the tradeoffs, and what you'd implement if you had more time.
This was my first time doing this kind of OA and I'm not familiar with the playbook, so I feel like I didn't hit their sweet spot. What I found was that when the input message is vague, the category the classifier returns is definitely wrong and the confidence is negative too, and in that case it shouldn't return anything.
I built my refactor around that, separating the cases that go to manual review from the cases where the LLM is called to decide. I wrote all the code by prompting the AI Assistant.
I got a rejection a day later. Not surprising, I knew I hadn't done well.
Can everyone share what the right playbook for this kind of OA is?
Discussion
Loading comments…