The problems worth writing up

Building the layer that lets LLM-driven agents run unattended against customer cloud infrastructure, and be trusted with what comes back.

99.9% execution reliability across millions of daily evaluations, failure rate under 0.1%

A lead distribution platform holding 230M+ active records across 11 related tables, where a 15 minute query had to become a millisecond one without a schema migration.

Query time from 15 minutes to milliseconds, a 99.99% reduction, across 2.5B+ records

A serverless platform whose MongoDB connections scaled perfectly in lockstep with traffic, straight towards the ceiling. Connections went from 1,800 to 400.

78% fewer database connections, 45% faster p95, 99.9% availability after re-architecture

Inheriting a fragmented Lambda estate across five legacy systems at a digital health company, and turning it into something a ten-person pod could actually operate.

40% infrastructure cost reduction, 65% less operational surface area, zero production incidents

A production database in an unrecoverable state and a business stopped behind it. Rebuilding InnoDB tablespaces, then making sure it could not happen the same way twice.

Operations restored, automated backups and version stability put in place afterwards