Why they built it
The founders came from four years building ML infrastructure at Cruise and saw data scientists stuck moving from local Python notebooks to dependable cloud pipelines. They built Sematic as an open-source orchestrator to run Python-defined work locally or on Kubernetes with traceability for artifacts, datasets, metrics and reproducibility. By November 2023 they shifted that foundation to Airtrain, a no-code platform for evaluating and fine-tuning language models on a customer's own data, on the thesis that “evaluation is the gate.”
Why it didn't work out
Airtrain did not end with a public failure explanation; by April 1, 2025 its homepage said the team was joining Weights & Biases and that Airtrain's products would be sunset. The notice did not describe an acquisition, disclose price or terms, or give a reason, so the record supports a product shutdown and team transition rather than a confirmed acquisition.
What went wrong
Most important first.
- 1
Independent products sunset with team move Other
By April 1, 2025 the homepage announced the team was joining Weights & Biases and the products would sunset, without stating transaction terms or a reason.
- 2
Converging evaluation incumbents Competition
Humanloop, Braintrust, and Weights & Biases itself overlapped on datasets, playground comparison, LLM-as-judge scoring, dataset registries and evaluation while those capabilities were becoming features of broader platforms.
- 3
Only seed financing on record Capital
The verified capital record is the $3 million seed round announced November 17, 2022, with no later financing announcement surfacing.
- 4
No validated Airtrain customer base Product-market fit
Voxel is the only named customer and was for Sematic, while the claimed thousands of AI professionals had no dated count or methodology and the Playground's third-place Product Hunt finish with 236 points showed launch interest, not sustained usage.
What could be different today
- Humanloop's September 8, 2025 platform sunset following acquisition included structured export tooling, showing how evaluation histories can be migrated when a vendor disappears.
- Versioned datasets integrated with evaluations are now documented platform behavior, as in Braintrust's dataset guide, so a rebuild can link to external dataset versions instead of rebuilding storage.
- Experiment data export paths plus self-hosted evaluation, annotation queues and optional governance features are now exposed, as in Langfuse's data platform overview and clarification, allowing a metadata-only record to complement existing runners rather than re-implement comparison, judging, clustering and fine-tuning.
