Ocular AI, a Silicon Valley-based company specializing in frontier AI training and evaluation data, has announced a US$2 million Pre-Seed funding round led by Drive Capital. The round includes participation from Y Combinator, Alumni Ventures, 1745 Ventures, Orange Collective, MyAsia VC, and a group of angel investors. The company reports seven-figure revenue and operates an expert network of thousands of vetted domain experts.
Voice AI Moves From Transcription to Conversation
Voice has become a priority for frontier AI labs because it carries tone, timing, hesitation, and emphasis that text transcripts lose. According to Ocular AI, NVIDIA, Thinking Machines Lab, OpenAI, and Google have each shipped or previewed audio-native models that listen and speak at the same time. These voice-native systems skip the transcript, allowing one model to hear tone directly, manage turn-taking, and respond to interruptions in real time.
The Bottleneck Is Frontier Data, Not More Data
Ocular AI argues that the core bottleneck is not the volume of speech data but the availability of frontier data at a fidelity models can learn from. The company works with leading labs to collect real conversation with each speaker on a clean channel, build evaluations that expose model failures, and ship datasets that address those failures. It also points to research showing that many open full-duplex models still rely on an 8 kHz telephone corpus recorded in 2004.
Converse Benchmarks Target Real Conversation
The company is expanding its evaluation work with Converse, a benchmark family built around the question of whether AI can converse like a human. The first benchmark, Converse-STT, measures how accurately frontier speech-to-text models transcribe real-world conversational data, including restarts, overlaps, hesitations, and accents. Converse will also assess speaking, conversing, and achieving outcomes, using real human recordings rather than clean read speech or saturated tests.
Beyond Voice and Toward Multimodal Data
Ocular AI said its infrastructure is not limited to voice and can support data for professional domains including medicine, law, finance, and software engineering. The company is also working with organizations building audiovisual models that combine speech and vision in a single stream. Its data pipeline is designed to provide provenance, consent, and licensing records for every clip and label.
Building a Research Lab for Applied AI Data
Ocular AI describes its next phase as building an applied AI data research lab. Every dataset begins with an evaluation that shows where a model fails and is built by credentialed experts who know what correct outputs look like. The company's approach combines domain expertise with rigorous research to produce training data that pushes frontier models past their failures.
Team and Next Steps
The company is based in San Francisco and has a team with backgrounds from Dartmouth College, the Indian Institutes of Technology, Microsoft, Google, and Atlassian. Ocular AI plans to use the Pre-Seed funding to expand evaluation suites, deepen research, and scale its expert network. It is hiring engineers and domain experts who want to shape the data layer for frontier AI.
With its US$2 million Pre-Seed round, Ocular AI is positioning itself at the intersection of human expertise and frontier model development. The company serves frontier AI labs and Fortune 100 enterprises that train and evaluate on its datasets, evaluations, and benchmarks. As voice-native and multimodal systems become central to AI, the demand for high-fidelity, expert-built data is likely to keep rising.