The Silicon Roundabout Podcast

The episode contains genuinely substantive technical methodology - the population-baseline approach to defining antisocial play time, geolocation grid-mapping, and the C&AVX optimisation story are non-obvious insights a data practitioner would learn from. However, these are heavily diluted by host cheerleading, praise sequences, and off-topic chemistry commentary that eat significant airtime.

“what we do is basically take all the players per time zone...you look at the distribution per time slot, like 15 minutes time slot, for example”

“the level of problematic gambling of a player over time fluctuates. So when someone self excludes the high risk of activity might not be just before the self exclusion”

The reframing of 'night play' into a population-relative 'anti-social time' using per-timezone activity distributions rather than fixed clock thresholds is a genuinely non-obvious, first-principles approach. The point about self-exclusion being a lagging and noisy label for supervised learning - and the strategy of building a proprietary human-reviewed dataset to overcome it - is also fresh and specific to the domain.

“instead of nighttime we call it anti social time”

“if the first version you had like um, 0.5 of our customers approach, we're looking into increasing this to 2% so even lowering the risk, so accepting more people, but in an automated intervention”

Karim and Maris are genuine practitioners who built and own this system at Kindred - Maris drives the domain specification and Karim leads the technical implementation. They are credible mid-level operators with real depth, though neither is a senior industry leader, and the episode is partly a recruitment pitch rather than a pure knowledge-sharing exercise.

“we went from what, three, four features in a model into 24”

“we collaborate with different data teams. Uh, we are dedicated to leco. We have also, um, a global data science team based in London”

The episode offers concrete numbers - 15-minute time slots, 5% activity thresholds, movement from 0.5% to 2% customer intervention rate, 20 - 25 minutes down to under 10 seconds for a million logins, 24 features vs original 3 - 4, and approximately one year of proprietary labelled data. Business-impact outcomes (e.g. harm reduction rates, model accuracy metrics) are notably absent, capping the score.

“for a million logins, it was taking less than 10 seconds”

“if the first version you had like um, 0.5 of our customers approach, we're looking into increasing this to 2%”

The host occasionally asks good follow-up questions that unlock real technical detail, such as pushing for a walkthrough of the night-play scenario. However, the conversation is repeatedly derailed by effusive praise of the guests, team-chemistry commentary, and recruitment messaging, leaving many technical threads underpursued and no meaningful pushback on any claim.

“Give us a walkthrough. This is very interesting. I understand how complicated this sounds. Absolutely. I can only imagine, you know, the work that's been put into it”

“I absolutely love the chemistry between you guys. I've always thought that the chemistry between colleagues is, you know, a huge part, if not the main part, as to why work, uh, is delivered well”