Why we built TQL instead of just teaching everyone SQL
ClickHouse and Postgres have different aggregation models, different date functions, different everything. Here's why one query language beat two dialects.
Jun 12, 2026·7 min read
BlogNotes on running
Notes on running
AI in production.
Engineering write-ups, case studies, and product thinking from the team building Trasys — mostly about the failure modes nobody warns you about.
The three failure modes that never trigger an alert
Cost drift, latency creep, and quiet hallucination spikes rarely cross a threshold. This is the case for anomaly detection that doesn't need one configured.
May 28, 2026·9 min read
How a 40-person team cut MTTR by 83% without hiring an SRE
A walkthrough of the incident workflow — alert rule, escalation, Debugger — that took response time from over an hour to under ten minutes.
May 14, 2026·6 min read
Dev Mode vs Normal Mode: designing one screen for two audiences
Why every dashboard in Trasys ships with a plain-English view and a raw-data view, and how we decided what belongs in each.
Apr 30, 2026·5 min read
Tracing a request across Kafka, ClickHouse, and three services
The waterfall model behind Traces, and how span-level timing surfaces the one hop actually responsible for a slow request.
Apr 9, 2026·8 min read
What we learned tracking cost across four LLM providers
Groq, Anthropic, OpenAI, and Google each report cost differently. Here's how AI Monitoring normalizes it into one number your finance team can read.
Mar 22, 2026·6 min read