Zurich-based AI infrastructure startup Rapidata has raised $8.5 million in seed funding as it pitches a faster way for AI developers to gather the human judgments used to train and fine-tune models, in a process that has become a growing choke point for the industry.
The round was co-led by Canaan Partners and IA Ventures, with participation from Acequia Capital and BlueYard, the company said this week.
Rapidata says it collects large-scale human feedback by distributing short, opt-in tasks through digital advertising and consumer apps, allowing AI teams to request targeted evaluations on demand rather than relying solely on fixed annotation workforces or one-off survey panels.
The company positions the approach as particularly useful for reinforcement learning from human feedback (RLHF) and other model-evaluation work where developers need preference rankings, quality checks, and “does this feel right?” judgments that are difficult to automate.
Chief executive Jason Corkill said the goal is to make human judgment available at near real time and at a global scale, so teams can run frequent feedback loops while models are being iterated.
Customers cited by the company include voice AI startup Rime and human-motion model developer Uthana, which said the platform helped them test outputs with real users more quickly than traditional vendor-and-survey approaches.
The funding will be used to expand Rapidata’s participant network and meet demand from AI companies seeking faster validation and data-labeling turnaround as competition intensifies, it said.
Rapidata is entering a crowded data-labeling and evaluation market, but its bet is that “human feedback” is shifting from a back-office annotation function into a continuous product-development loop, especially for generative AI systems where quality is subjective and context-dependent.
If it can reliably recruit and route the right respondents (by language, domain and demographics) without sacrificing data quality, the model could appeal to teams that are already constrained less by compute than by the speed of evaluation and iteration.
