Silver.dev

2nd engineer at LatAm's fastest-growing recruiting agency.

Context

Silver.dev is Latin America's fastest-growing recruiting company, connecting software engineers with U.S. startups.

After setting their live coding speed record, being #1 out of around 440 senior engineers, I joined them at around the same time I started my senior year of high school. For that reason, during the first nine months I worked part-time; after that, I worked full-time for 3 months until I left because I was eager to build Just Ralph It.

I used to work day-to-day next to Paske (founding engineer), and many days I would receive feedback from and discuss what to do with Gabriel (founder). It was my first job where I worked alongside a team, so I learned a lot about work dynamics there.

Projects

Skipping minor issues resolved along the way, in no particular order, I did this:

  • Silver.dev: I rewrote and then maintained the entire public company website.
  • Candidate Portal: I built various end-to-end features on our main engineering project, used by all the candidates. For example, one feature involved reverse-engineering LeetCode's API so candidates could sync their progress with the portal, and then designing a good-looking card to display the data.
  • Recruiter's Extension: A Chrome extension that monkey-patched LinkedIn and Ashby (our ATS) to display more data about candidates taken from our database. All recruiters plus the founder used it daily. I designed and built the entire project myself.
  • Open Silver: I was in charge of our collection of open source tools, including:

Incident anecdote

Turns out, I deleted the LinkedIn URLs of 20,000 candidates. And turns out, the backup was also broken! I just fucked the entire database of the company I had recently joined as an intern.

In one of my first months, a task I had was to update the links associated with every candidate. Particularly, Ashby (our ATS) has a socialLinks field to save links such as GitHub, LinkedIn, personal website, etc. We wanted to use that feature to add one more link that pointed to what we called the "candidate card," a pseudonymous web page displaying information we had on our separate database about the candidate.

I was writing a script to do such a bulk operation, and I finished writing it by that Tuesday evening. I ran it against my own candidate profile, it seemed to work fine, and so then I left it running for the rest of the candidate profiles. Given that the Ashby API didn't support updating social links for multiple candidates in bulk, I had to update them one by one, so I had to leave the script running overnight.

A few moments later, Gabriel and Paske came to talk. My computer was there, with the script visibly updating candidate after candidate. They asked what I was doing, and they nodded along.

The following day, Wednesday, was the remote one of the week. I had just finished lunch; I was doing some work when one recruiter reported having issues when using Ashby's Chrome extension. "Completely unrelated to me," I thought, so I ignored it.

Later, Paske asked me to go on a call as soon as possible. What could be going on? Turns out, the Ashby Chrome extension works by using the LinkedIn URL from the candidates. Turns out, setting the socialLinks field overwrites all fields, rather than just appending new ones. Turns out, I had just deleted the LinkedIn URLs of 20,000 candidates overnight. And turns out, the backup was also broken!

I just fucked the entire database of the company I had recently joined as an intern.

Along with Paske, we started looking for ways to amend this incident. While doing that, he explained to me the importance of staying calm, so that we could think clearly and avoid worsening the situation further. Our backup, which was a sync we had with Ashby's data, had the socialLinks field. However, the array was empty for every candidate; only that field specifically was not syncing properly! Luckily, we later found another field in a separate table containing the candidates' applications as JSON blobs, which included the submitted LinkedIn URL. Although not all, that allowed us to recover most of the lost URLs. Phew!

On the technical side, the main lesson was about when to run and how to properly test this kind of dangerous script: don't do it when you're tired in the evening; have someone else review it if you're inexperienced; and test it on a gradually higher number of records, rather than all at once, and be 110% sure it works as expected before doing that. On the non-technical side, it was what Paske taught me about not rushing when there's an incident.