field notes
Insights
War stories from production data work. What broke, what we measured, and what we'd tell you over coffee before you build it yourself.
call centers · caller id · automation · measurement2026-09-22
The Number Your Lead Dials From Is a Decision, Not a Setting
Carriers grade your outbound numbers on every call. Most floors set a caller ID per campaign and find out monthly. Here is what changes when the choice is made per lead, every fifteen minutes, from your own dial log.
read →
field sales · gps · door to door · product design2026-09-22
How GPS Door-Knock Verification Works: Two Radii, Two Jobs
A door-to-door sales app asks GPS two different questions: which house is this, and does this knock count. Our field sales app uses 60 yards for the first and 250 for the second. Here is why one number cannot do both jobs.
read →
field sales · mobile · reliability · engineering2026-09-22
Patchy Signal on the Porch: How a Field Sales App Should Queue a Visit
Reps knock doors where the signal drops. Here is how our field sales app keeps a visit that cannot leave the phone, says so out loud, sends it in order when the signal is back, and never counts the same door twice.
read →
ai · call centers · quality assurance2026-08-16
Your QA Team Hears 3 Percent of Your Calls
Call centers assure quality by sampling 1 to 3 percent of calls. AI review changes the denominator: every call scored, every score backed by a quoted line.
read →
engineering · data platforms · reliability2026-08-15
Production-Grade Is a Checklist, Not a Compliment
What separates a system a business runs on from a demo that works when everyone behaves: invariants in the database, money in fixed precision, and failure that announces itself.
read →
machine learning · insurance · explainability2026-08-14
We Chose Logistic Regression on Purpose
For an insurance fraud detection system, we picked the boring model: ~0.90 accuracy, ~0.80 AUROC, and every flag explainable to the person who has to act on it.
read →
data quality · analytics · call centers2026-08-10
How a Bad Join Created a 5x Reporting Error
A call center believed carriers rejected 61 percent of its calls. The real number was 11 to 12. The culprit: one join that silently dropped most rows.
read →
machine learning · causal measurement · e-commerce2026-08-09
Measuring Marketing Lift With a Permanent Control Group
Retention tools claim credit for revenue that would have arrived anyway. The fix is old-fashioned science: a permanent randomized holdout that is never emailed.
read →
llm · nlp · hospitality2026-08-08
What It Takes to Label 67,962 Reviews With an LLM
Sentence-level labeling across an 84-aspect ontology, 26 ontology versions, and a model bake-off that kept 95 percent of quality at a quarter of the cost.
read →