Production Monitoring for AI agents. Backed by Y Combinator.
About the role
Lemma reads all of your agent conversations, finds the failures nobody knew to look for, and opens the PR that fixes them. Right now it works because it's small and we're watching it. This role is about making it work when it's large and nobody is watching it.
What they're looking for
- 4+ years on backend or data infrastructure, with something high-volume in your history you can talk about in real detail
- Strong TypeScript. The whole codebase is TypeScript, including the core service and workflows
- Comfort with columnar stores and the economics of storing a lot of events cheaply. We run ClickHouse, Quickwit, and Qdrant, and you'll have opinions about all three
- Pragmatism about scale. We need architecture that survives 100x, built by someone who won't build for 100x on day one
- Bonus: anomaly detection work where the ground truth was genuinely unknown
- Bonus: you've built the enterprise deployment path before. Self-hosted, VPC, air-gapped
More about this role
Lemma reads all of your agent conversations, finds the failures nobody knew to look for, and opens the PR that fixes them.
We sell reliability tooling. There is no version of this company where our own pipeline is the flaky part.
Right now it works because it's small and we're watching it. This role is about making it work when it's large and nobody is watching it.
The interesting problems here aren't modeling problems. Agent runs go for hours, call thousands of tools, and produce deeply nested spans that look completely different at every customer. We read all of production, not a sample, and we surface the anomalous fraction of a percent without anyone telling us what anomalous means. Doing that accurately is hard. Doing it at a cost per event that doesn't eat the business is the actual job.
Then the part that decides whether we have a company: reproducing a failure we saw once, proving it's real and not variance, and being right enough that a team lets us open PRs against their repo. Being confidently wrong once costs more trust than being right fifty times earns.
- Own ingest: schema, throughput, cost per event, and the long tail of customers whose instrumentation is a mess
-...
Browse similar: AI jobs · AI startup jobs · Startup jobs · San Francisco Bay Area