The record
Impact
Every figure is a real production result, a before and an after I can point to. Some link to the full story; more get written over time.
Fig. A
Traffic at scale
2M visitors→100M req/mosteady on AWS
Ran a platform serving 2M visitors and nearly 100M requests a month on AWS, holding steady under load.
Fig. B
Query collapse
2,000,000→2408,000× fewer
A full product feed sync fired two million database queries and ran for days. Rewrote it as batched reads, about 240 queries, and the sync finishes in minutes.
Read the storyFig. C
DDoS absorbed
7M req/min→0 downtimeattack blocked
A DDoS attack hit 7M requests in a single minute. Held the edge and blocked it with no downtime for real users.
Fig. D
Fraud screening
manual review→auto-screenedKYC + liveness
Built a custom security rule builder and fraud detector for a large e-commerce checkout covering cards, Apple Pay, Google Pay, PayPal and crypto. Flagged activity triggers KYC with liveness and ID verification.
Read the storyFig. E
Chargeback evidence
manual hours→one lookup10 sections auto
Fighting a chargeback meant building a long evidence document by hand. Built a tool that fills a 10-section packet from order data and live gateway records, with card scheme liability rules coded in.
Read the storyFig. F
OTP brute force
1M combos→50 guesses/dayattack capped
A 6-digit login code has a million combinations. Three stacked rate limits cap an attacker at 50 guesses a day, about a 0.005% chance against any code.
Read the storyFig. G
Multi-site + SSO
one store→N storefrontsone codebase
Launched a second storefront on the same codebase, database and cache with strict per-site isolation. Hub-and-spoke SSO signs customers in across every site and one logout ends the session everywhere. More sites plug in with no extra work.
Fig. H
Report build
60s→5s12× faster
Reports on a 4 GB MariaDB hit one-minute timeouts. Moved analytics to ClickHouse; they return in five seconds.
Read the storyFig. I
Order routing
manual→60% autohands-off
Order shipping was fully manual. Automated the routing pipeline so most orders now ship without a human in the loop.
Fig. J
Database CPU at peak
99%→45%zero alerts
Peak database CPU sat near 99% and paged on-call. Cut sustained peak load to 45%, with no alerts since.
Fig. K
Schema normalization
one wide table→keyed tablesone fact once
Wide tables repeat the same fact across many rows, so one change means many writes and drifting copies. Normalizing to 3NF stores each fact once and joins do the rest.
Read the storyFig. L
Live DB changes
high stakes→routinetrusted hands
Schema changes against live traffic are where platforms go down. On a ~100M req/mo platform, I plan and run migrations and data fixes as routine work, not emergencies.