01 — about
We were the customer first
Cinder started as an internal tool at a company whose agents kept failing in ways their monitoring could not see. The tool outlasted the company.
In 2023 we were running a document-processing product with an agent at the centre of it. It worked, mostly. When it did not work, the failure almost never came with an error — a retrieval step returned nothing, a tool call was truncated mid-argument, a router sent an expensive request to the cheap model. Our dashboards were green throughout.
We built the first version of what became Cinder to answer one question: why did this run differ from the ones that worked? It took a fortnight and it found three bugs in the first week, two of which had been shipping for months.
We spent the next year trying to buy something better and failing. The APM vendors treated a model call as an HTTP request. The LLM-specific tools logged prompts and called it observability. Nothing recorded the shape of a run, which is where the information actually is.
So we sold the document product and built this one instead. Cinder Systems was incorporated in 2024. There are fourteen of us.
02 — principles
Three sentences we argue about
These are load-bearing. Each one has cost us a feature we wanted to ship.
Show the evidence, not a verdict
A single confidence score people over-trust is worse than a trace they have to read. We surface structure and let engineers draw the conclusion, because they are better at it than we are.
Be honest about what is missing
Partial traces are labelled partial. Sampled data says so. A rubric publishes its human-agreement number next to its score. The failure mode we most want to avoid is confident and wrong.
Your data leaves whenever you want
Scheduled Parquet exports to your own bucket, on every paid tier, in a documented schema. Lock-in earned by being hard to leave is not a business we want to be in.
03 — team
Who you would be emailing
Priya Raman
Co-founder & CTO
Spent six years on the tracing team at a database company, then two more discovering that none of it transferred cleanly to agents.
Marcus Feld
Co-founder & CEO
Was the customer before he was the founder — ran platform engineering at a company that tried and failed to build this internally.
Mei-Lin Cho
Head of Research
Works on evaluation methodology, and on convincing people that a number without an error bar is not a measurement.
Daniel Okafor
Staff Engineer, Ingest
Owns the path from your process to our storage. Has strong opinions about sampling and will share them unprompted.
Rosa Amsel
Design Lead
Draws trace waterfalls until they stop needing a legend. Previously designed monitoring tools nobody enjoyed using, which she took personally.
Tom Bergqvist
Developer Relations
Writes the SDK docs and the sample apps, and files the bugs he finds while writing them.
Come and break it.
The free tier is the same product we run ourselves. Tell us where it falls short.
npx @cinder/cli initOr talk to an engineer — the same people who wrote the ingest path.