Four cleared outcomes. Nothing else.

The platform was paying for warehouse time it was not converting into useful latency. Cost looked fixed because nobody had query-level visibility tied to product workloads.

I traced where spend and latency actually came from, down to workload patterns and table design. The fix list focused on the highest-impact changes first, then sequencing.

Headline outcome: monthly warehouse spend cut 44% while keeping 5TB+ queried sub-second, with 20GB/day telemetry and 1M+ monthly ingests in production.

The team had steady growth and steady usage, but compute spend was climbing faster than both. Cost controls were reactive and late.

I rebuilt the warehouse policy around workload isolation, right-sized compute tiers, and execution patterns that matched demand instead of peaks.

Headline outcome: annual Snowflake compute moved from $80K to $50K.

The platform depended on a full reload pattern that made runtime and cost unpredictable as data volume climbed. Jobs kept growing, and so did monthly compute.

I replaced full reloads with delta loads and rebuilt the processing flow around incremental behavior and operational guardrails.

Headline outcome: runtime moved from 8 hours to 47 minutes, and monthly compute moved from $14K to $5.8K.

A threat-intelligence catalog was assembled by hand by three people. Throughput and consistency were limited by manual work, and there was no reliable audit path for source-to-output provenance.

I replaced the manual process with a Rust and Python pipeline that ingested and normalized 20+ sources, then ran 20,000+ records through Vertex AI batch inference.

Headline outcome: manual assembly was replaced with a production data layer and an audit ledger with SHA-256 provenance.