Delphi

How Delphi keeps every Digital Mind answering

Delphi runs on Vercel, AWS, PlanetScale, ClickHouse and Upstash. It uses Polylane to triage its alerts and fix what breaks.

11.6K alerts triaged in its first two weeks
679 issues surfaced in its first two weeks
110 Polylane pull requests merged in its first two weeks
19 min median from detection to a fix pull request in its first two weeks

Connected to Polylane

The challenge

A Digital Mind is only as good as its next answer. Each conversation on Delphi pulls from a creator's content, calls a model and streams the reply back as text or speech. A slow retrieval, a failed model call or a dropped voice session breaks the conversation mid-sentence.

Those conversations run on the creators' schedule, not Delphi's: one newsletter or podcast mention can bring a wave of new people in an hour. At that scale, alerts pile up faster than anyone can read them.

“On a busy day we get more than a thousand alerts, and nobody on the team can read that many. Polylane reads all of them, finds the few that matter, and opens the fix. We connected the rest of our stack on a Monday and were merging its fixes the next morning.”
Spencer Schoeben Founding Staff Software Engineer at Delphi

Why Polylane

Delphi connected Polylane to Vercel, PlanetScale and GitHub in August. In November it added AWS, ClickHouse, Upstash, Logfire and PostHog, so Polylane's context graph now covers every system a conversation touches. Polylane learns what normal looks like for each of them, raises an issue with the evidence when something drifts, investigates across code, deploys and telemetry, and opens the fix when the cause is in Delphi's code.

How Delphi uses it

Alerts go to Polylane first. It folds the duplicates and the noise and raises the issues worth a look in Slack and Linear. Every fix goes through Delphi's normal review and CI.

0 50 100 150 200 2 Nov 10 Nov 110
Delphi merged 110 Polylane fixes in two weeks
Running total of Polylane fixes merged at Delphi, by day, from the day it connected AWS, ClickHouse and Upstash.

The results

In its first two weeks with the full stack connected, Polylane triaged 11,609 alerts down to 679 issues and opened 153 fixes, with a median of 19 minutes from detection to a pull request. Delphi merged 110 of them (72%).

Nobody should be on-call.

Related customer stories