Section 13: What the Panel Actually Wrote Down - Final Summary & Candidate Scorecard

Executive Summary: This closing section turns the module into a scoring tool. We replay the interview as a sixty-second recap, map each phase to the signal it recorded, and lay out the 100-point scorecard: six dimensions, each with the phrases that earn and lose points on this problem. Score bands say what to practise next, and a short catalog shows where nearby search hides under other names.

The hour felt like it was about geohashes, but geo-indexing is worth twenty points out of a hundred. The rest came from whether "near" became a number early, whether the 23,600 places in Sam's covering changed what you built, and whether you stayed calm at "just sort by distance." You can't reliably earn points you can't name, so this section names them.

The Sixty-Second Close

With five minutes left, the interviewer usually asks for a recap. Don't recite boxes. Retell the story through its decisions.

Interviewer: "Give me the whole design in a minute."

You: "Sam is in Bryant Park, cell dr5ru6, searching 'ramen', open now, within 1 km. Logs only ever see geohash5. The result cache is keyed by the geohash6 cell, query, filters and sort, and every hit is re-filtered from the caller's own point. On a miss we build a covering from the circle's bounding box: 12 p6 cells, not the center plus 8 neighbours, which would miss part of the circle north to south. Because we shard by geography, that's one shard of 64. Postings intersected with 'ramen' and open-now give 70, then 52, haversine leaves 24, and two-stage ranking returns 20. Menya Kaze beats Noodle Lab because the Bayesian rating turns 5.0 from 3 reviews into 3.96. That's about 120 ms against a 150 ms p95. Review aggregates are already in the index docs, so we never make 348K Reviews calls a second. Owner edits reach search through CDC, Kafka and versioned indexers in under 5 minutes. Three active regions are each sized for 10,000 searches a second, so losing the Americas leaves the other two at 8,700."

Nearly every clause holds a decision plus the number that justifies it. That is what tells the panel the earlier reasoning was real.

Quick check before moving on: