Work
Systems that survived their growth.
A few recent engagements we can talk about. Most of our work stays under NDA, but the patterns are the same.
Sub-second matching engine
Challenge: order latency spiked during volatility, causing slippage and dropped connections.
What we did: reworked the network path between matching nodes and market-data feeds, added kernel-bypass NIC tuning and cut retry storms.
Outcome: p99 latency dropped from 340 ms to 45 ms; connection drops fell by 92%.
Global edge in six regions
Challenge: viewers in Asia and South America saw buffering while origin stayed healthy.
What we did: built a tiered edge layer with cache hierarchy, origin shield and per-ISP TTL tuning.
Outcome: time-to-first-byte improved by 40%; origin egress dropped by 60%.
Telemetry for 12,000 devices
Challenge: devices produced bursts of data that overwhelmed the ingestion pipeline.
What we did: designed buffering, backpressure and regional aggregation nodes with failover.
Outcome: 99.99% ingestion reliability; 70% reduction in compute during peak bursts.