Hacker News new | ask | show | jobs
by 0x706B 1242 days ago
Had to launch our AI product next day else customer would jump off and this would've basically killed the company. But AI was far from perfect and did only work for very small amount of usecases. Exchanged the AI with a queue and UI so humans could do the task instead in realtime. Worked. Went live. Completly smashed competitors.

--

Just at launchday our internal infrastructure was fucked pretty bad and we faced a severe outage on everything non-productive. Product failed at some place to do proper error handling and basically got stuck forever since it couldn't reach some monitoring endpoint we had still placed on our int infra. We couldn't build a fix, since basically CI/Signing was part of the outage. ETA to get infra back up was something like ~8h. Panic grew as it would've totally fucked our go live. Hacked together a dummy http server returning status code 200 and exchanged DNS entry to point to that instead of int infra. Worked. Went live.

--

Had to do an audit before we could go live, due to deadlines and planned vacation of auditor it needed to be spot on, no time for touch ups and re-audit. During the audit, he asked for a thing we forgot to do (some security alerting on a 3rd party tool which didnt support it), told him something along "sure, love to show that, yadda yadda, lets have lunch first." went to the most chatty guy in the office to join us for lunch (so he wastes some time), faked a "oh, shit, forgot i have a sync call, will catch you up there". While they were having lunch, I checked out the 3rd party code, added the alerting, wasted 20 mins with their broken buildsys. Managed to build the one lib necessary for the alerts, but not the whole image. Overrode the lib in the original img instead and hoped shit would work. They were back too early, couldnt test, just fired up the deployment, wasted some more time babbling with the auditor till it went through. We live tested it together. Worked. Audit went through without any remarks and we hit launchday.

2 comments

Ah yes, the classic AI startup with a team full of humans actually pulling levers behind the scenes.
Fun fact: biggest learning was that this is perfectly fine. It's a $task startup solving $task for you.

$task means pulling 100 levers to the right level. Every lever has different precision requirements.

20 levers you can pull good enough with simple if/else and call it a day 20 more need a bit of statistics voodoo to reach the precision requirements 20 more can be pulled with AI to the approx. right position right away. 30 more could be pulled by AI in the future after going live when there's enough data on where the levers need to be pulled to. And the last 10 levers need such high precision that humans need to be involved (maybe for the moment, maybe for ever)

As long as solving $task is profitable enough and latency & error rate by humans low enough, there is nothing wrong with it.

Some competitors tried it with pure AI, but didn't hit the precision requirements and we could take over their customers. Others burned through all their funding trying to get the AI right before going live.

Oh yeah, nothing wrong with it at all. Plus, you need the human team to be creating the training data set.
third one is almost like in james bond movies when you barely make it but you succeed etc :D