I've noticed that the actual modeling often gets more attention than the boring part beforehand: figuring out what the columns mean, missing values, duplicates, weird categorical values, outliers, leakage, domain aware data cleaning etc.
Curious what other people here actually spend the most time fixing before they can start modeling.
Ran a lifetime campaign to get some fast cash and early validation. made decent money upfront but then one time buyers completely took over our support inbox.
They kept demanding custom features for months which distracted us from building a proper monthly subscription base. Ended up having to issue refunds and cut ties just to get our time back.
Has anyone else done an LTD and instantly regretted it once the support tickets started piling up?
Back in 2013, we co-founded one of the earliest real-time spot market TMS platforms. The platform itself worked, but we hit a massive operational wall: drivers flat-out refused to download another native app.
That driver resistance forced us to pivot and build an appless tracking system instead, which solved the onboarding bottleneck overnight.
Microsoft is reportedly working on making Copilot more emotionally intelligent, designed to check in on users and behave more like a therapist to resonate with Gen Z.
We tracked settings usage across 12,000 accounts over 6 months.
Never opened settings: 81% Opened once: 14% Opened more than once: 5% Most-changed setting: notification frequency (2.1% of users) Second: theme (1.8%) Third: timezone (0.9%)
We had built 47 settings. The median user changed zero. The product was effectively one configuration for almost everyone, and we had spent years adding knobs that 95% of users never touched.
The lesson was not "remove settings." It was that defaults are the product. Every setting we added was a decision we were avoiding making for the user.
We analyzed 9,300 support conversations over 8 months. Median first-response time: 4h 20m.
But 68% of tickets contained a follow-up from the user that included the answer a screenshot, an error code, a version number, or a clarification that made the original question answerable. The user had solved it themselves, or had found the missing detail, while waiting.
Median time between the user's first message and their second: 2h 10m. In other words, the user was doing our triage for us, slowly, because we asked too late.
We added a short intake form with three required fields: what you did, what you expected, what happened. Tickets dropped 18%. First-response time dropped to 1h 40m. The conversations that remained were more answerable on the first try.