Airtrain AI
Summer 2022InactiveNo-code data curation for LLM fine-tuning and evaluation.
Pressure-test this opportunity
Explore the risks and possibilities with a prompt for ChatGPT, Claude, or your agent.
No-code data curation for LLM fine-tuning and evaluation.
Explore the risks and possibilities with a prompt for ChatGPT, Claude, or your agent.
Airtrain AI was a San Francisco developer-tools company founded in 2022 by Emmanuel Turlay and accepted into Y Combinator's Summer 2022 batch. It began as Sematic, an open-source continuous machine-learning orchestrator, then shifted by November 2023 into a no-code platform for evaluating, fine-tuning, and organizing language-model data. YC now marks the company inactive.[1]
The final homepage, captured on April 1, 2025, said the team was joining Weights & Biases and that Airtrain's products would be sunset.[2] It did not explain why, describe an acquisition, or disclose a transaction structure, price, or terms. The evidence therefore supports a product shutdown and team transition, not an acquisition claim or a specific failure theory.
Turlay brought an unusually infrastructure-heavy background: particle-physics research at CERN, order and payment systems at Instacart, and machine-learning platform leadership at Cruise.[1] Sematic's founding team also included engineers Josh Bauer, Tudor Scurtu, Chance An, and Sash Nagarkar.[3] The team said its thesis grew from four years building ML infrastructure at Cruise.
Sematic addressed the gap between a local Python notebook and a dependable cloud pipeline. Its open-source framework let data scientists define work in Python, execute locally or on Kubernetes, and track artifacts, models, datasets, metrics, visualizations, and reproducibility metadata.[4] The original ambition was continuous machine learning with traceability, reproducibility, and observability.
On November 17, 2022, Sematic announced a $3 million seed round led by Race Capital, with participation from Y Combinator, Soma Capital, Leonis Capital, Fundament, and Pioneer Fund.[5] A contemporaneous release named Voxel as an early commercial customer and said the money would support hiring, a hosted cloud product, and open-source adoption.[6]
Sematic was the technical substrate. It ran Python-defined ML pipelines locally and on Kubernetes, preserved lineage and artifacts, and packaged cloud jobs. Airtrain reused that infrastructure for language-model workloads rather than abandoning it.
The first Airtrain experience compared open-source models on a customer's own data. Batch evaluations could contain up to 10,000 rows. Users selected models or supplied outputs, then evaluated them through LLM-assisted scoring, JSON-schema validation, standalone metrics, and benchmarks. Plain-English properties such as groundedness or politeness became judge criteria scored from one to ten, with distributions and row-level inspection for failure analysis.[12]
Image 1 / 2
The Playground compared open and proprietary models while exposing token counts, throughput, inference cost, and saved sessions. The broader platform added an embedding-based Dataset Explorer, semantic clustering, visualization, custom fine-tuning, and automatic classification and labeling from natural-language class definitions.[13][14]
Airtrain Pro bundled up to ten private datasets, clustering, visualization, comparison, and fine-tuning on Mistral 7B, Gemma 2, and Llama 3 base models.[15] Security-sensitive enterprises could deploy open-source compute in their own cloud and use a Python SDK.
Airtrain entered a fast-converging market. Humanloop linked datasets across evaluation and fine-tuning.[16] Braintrust offered playground comparison, versioned datasets, LLM-as-judge scoring, and production evaluation.[17] Weights & Biases itself overlapped through training, inference, dataset registries, and evaluation.[18]
Airtrain differentiated through no-code workflows, smaller open-model selection, evaluation on private task data, and a direct path from diagnosis to fine-tuning. Turlay's compact product thesis was that “evaluation is the gate.” The challenge was that evaluation, data management, tuning, and experiment tracking were becoming features of broader platforms as well as standalone categories.
Batch evaluation was free for datasets up to 10,000 examples. Fine-tuning used upfront pricing inputs such as model, dataset size, epochs, and token volume, but the archived page published no dollar amounts. That tier included managed infrastructure, model export, and dedicated Slack support. Enterprise packaging added private-cloud compute and a Python SDK.[9]
No reliable revenue, ARR, paid-customer count, margin, burn, or runway figures were found. The verified capital record is the $3 million seed round. No later financing announcement surfaced.
Voxel is the only named customer in the record, and it was identified for Sematic rather than Airtrain. The company later claimed that thousands of AI professionals used the platform, but provided no dated count, methodology, or named Airtrain customer.[19]
The Playground's 236 Product Hunt points and third-place daily finish are bounded evidence of launch interest, not sustained usage. Airtrain also published benchmark and labeling-cost claims, but no independent study or customer case study validated them. YC lists a team size of five.
The strongest conclusion is narrow: Airtrain's independent products ended, and the team announced it was joining Weights & Biases. The statement described AI safety, customization, and evaluation as expertise the groups would combine, but did not state that Weights & Biases acquired the company, assets, intellectual property, or contracts. Turlay later listed roles at Weights & Biases and CoreWeave, which is consistent with a team transition but does not establish transaction terms.[20]
There is no sourced explanation for the sunset. Competitive overlap, product-market fit, financing, model commoditization, or an acqui-hire may sound plausible, but none is demonstrated by the public record. Nor is there evidence that particular Airtrain code, features, datasets, or customers survived inside Weights & Biases. The last captured feature evidence dates to September 2024; the shutdown notice appeared by April 2025; the domain was for sale by July 2026.
What can be learned is structural. The team reused an orchestration foundation to enter an emerging LLM workflow, built a broad suite rapidly, and then moved into a larger platform with substantial feature overlap. The unresolved issue for users is not another evaluation runner. It is durable, vendor-neutral evidence of why a model passed a release or procurement gate when evaluators, judges, thresholds, and vendors change.