We make model behaviour auditable.
Tracewell was started by two SREs who got paged for an LLM feature and found there was nothing to look at. 31 people, Series A, 318 projects in production.
Software used to be deterministic. You wrote the code, you read the logs, you knew. Then we put a model in the loop and shipped it anyway. The model changes under you. The users ask things you never tested. The bill arrives before the bug report. We think the fix is boring and old: observe everything, test before release, gate what matters. Not a dashboard you check after. A receipt for every call and a check on every merge. Trust is not a feeling about a model. It is a number you can show the auditor.
Two years, five steps.
We publish our own numbers the way we ask customers to: dated, specific, no rounding up.
Mara and Tobias leave their SRE jobs after a 3 a.m. page for a chatbot nobody could debug. First prototype: a span exporter for OpenAI calls.
$3.2M seed. Ten design partners, among them Lumeo and Tidewater. First eval judges ship; first customer finds a looping agent on day two.
Regression gates in GitHub Checks. Kestrel passes its first audit with signed eval reports. Query engine rewritten for 1B+ spans.
$24M Series A. San Francisco office opens. Self-hosted edition GA in June; EU and US regions with zero-retention proxying.
2.4B spans a day, 318 projects, 31 people. Cost attribution and loop detection on every trace. Go, Java and Ruby SDKs in beta.
Three rules we actually enforce.
Numbers, dated.
Every claim on this site has a figure and a date. If we can't measure it, we don't say it. Marketing copy goes through the same review as code.
On-call for our own product.
Engineers carry the pager for the systems they build, and the console is instrumented with Tracewell. Our p95 is public in the footer.
Boring where it counts.
Postgres, object storage, Arrow, OpenTelemetry. We spend novelty on the product, not on the stack. Self-host works because the stack is ordinary.
Eight of thirty-one.
SREs, database people, a few who trained models and one who audited banks. Berlin, San Francisco and nine remote cities.
Mara Lindqvist
Co-founder, CEO · ex-SRE lead, Zalando
Tobias Reinholt
Co-founder, CTO · ex-Datadog, query engines
Anouk Verhoeven
Head of Evals · ex-research, NLP evaluation
Dev Ramaswamy
Staff Engineer, SDKs · ex-OpenTelemetry maintainer
Inês Carvalho
Design · console, waterfall, every hairline
Jonah Feld
Go-to-market · ex-engineer, talks to every customer
Yuki Tanabe
Research · judge calibration, drift models
Samuel Okafor
Solutions · ex-bank auditor, owns the signed reports
Five seats. Pager included.
Salary bands are in every posting. Interviews are two technical sessions and one with a customer. We reply to every application within five business days.
Come build the receipt.
Send a short note and a link to something you've shipped. A human reads it within five business days.