// stuff I actually shipped

Selected Work

View all →

// numbers that kept me employed

Things I Actually Improved

Real numbers from real servers that real people depend on

Vector searchNuclaDB
1.8xQdrant's throughput

ef=10 · recall@10 0.932 vs 0.959

NuclaDB13,914 QPS
Qdrant7,624 QPS

Vector database written from scratch in Go, benchmarked head to head against a real Qdrant instance.

Search latencyWayground
78%faster

10M+ records

beforebaseline
after<280ms

Partitioned the OpenSearch indexes. Was painful before.

Quiz latencyWayground
42%faster

1M+ quizzes a day

before2.1s
after1.2s

Redis pipelining and autoscaling tuning. Users didn't notice. That's the point.

Memory
85%less RAM
before1.2GB
after180MB

Redis pooling and chunked streaming. The server stopped sweating.

API gatewayVantageEdge
70Kreq/s

one instance, full pipeline

+0.5ms
p50 overhead
+1.5ms
p99 overhead
~2ms
config push

Auth, rate limiting, caching and load balancing, for less than a millisecond typical overhead.

Event ingestionWayground
50M+events/day

<30s end-to-end

Kafka
Pub/Sub
BigQuery

Analytics events streamed into BigQuery, queryable within half a minute of happening.