The company that cleans up after the AI boom - labeling the images, video, point clouds and text that models actually learn from.
Here is a thing about artificial intelligence that the keynote slides tend to skip: before a model can recognize a pedestrian, a tumor, or the sarcastic tone in a customer email, some human being has to point at examples and say this one, yes; that one, no. Millions of times. This is called data annotation, and it is the least glamorous, most necessary job in the entire field. Stardust AI, founded in 2017 and rooted in San Francisco, decided to become very good at it.
The pitch is deceptively simple. Everyone else is racing to build a bigger model. Stardust builds the thing that makes any model worth building - clean, correctly labeled, well-managed data. Its flagship platform, Rosetta, has processed more than a billion pieces of data. Its data engine, MorningStar, manages the whole messy lifecycle after that. And its CEO, Lei Zhang - a former World Bank and Wall Street quant who spent a decade in data and modeling before founding the company - likes to frame the stakes in a single sentence: in the era of AI 2.0, whoever controls the data controls the model.
That is a self-serving thing for a data company to say. It is also, inconveniently for the skeptics, true. A model trained on garbage produces garbage confidently, which is worse than producing nothing. Stardust's business is essentially the arbitrage between how much AI teams want to skip the data work and how much they can't afford to.
"Having control over your data means having control over your model."— Lei Zhang, Co-Founder & CEO
Stardust's product line reads like a small solar system, which is on-brand. What they share is a single idea: automate the repetitive labeling, and reserve human judgment for the hard, ambiguous 40% where models otherwise quietly fail.
The workhorse. Rosetta annotates multimodal 2D/3D/4D data - images, video, point clouds, audio, text - blending a 60%+ automatic processing rate with a human-feedback engine. Named, fittingly, after the stone that made an unreadable language readable.
Manages the full enterprise data lifecycle across production, insight and intelligence. Its stated enemy is "data debt" - the low-value data that piles up and slows every model iteration. A SaaS beta launched at WAIC 2024.
A four-tier stack of ready-made and custom datasets built specifically for large language model development and data trading - the pantry of pre-cleaned ingredients for teams that don't want to start from raw.
An instruction-following evaluation service that tests large language models across 150+ task variations. The distinction it measures is the one that matters in production: not "can it," but "does it, reliably."
The reason Stardust's niche is defensible is the same reason it's unfashionable: it's tedious, it requires compliance discipline, and it doesn't demo well. The company is aligned to GDPR, CCPA and ISO standards - words that make investors yawn and enterprise legal teams exhale. For a car company feeding an autonomous-driving model, or a bank training on customer records, that yawn-inducing compliance is precisely the product.
It also explains the client list. Autonomous driving is annotation-hungry in the extreme - every cyclist at dusk, every occluded curb, every point cloud needs a label - which is why names like SAIC Motor, BYD, Geely, Bosch and ZF show up. Stardust has served as a strategic AI annotation platform for JD.com and partnered with Baidu EasyDL.
The practical uses are unglamorous in the best way. An autonomous-driving team can hand Stardust raw sensor logs and get back labeled 3D point clouds and video. An LLM team can commission a custom dataset - or buy tiers off the COSMO pyramid - then benchmark whether the resulting model actually follows instructions. An enterprise sitting on years of unstructured images, audio and text can run MorningStar over it to find the valuable data, kill the "data debt," and shorten the loop between training and deployment. The common thread: you keep control of your own data, on cloud or on-premises, instead of renting it back from someone else.
Lei Zhang launches the company to unlock data value and democratize AI, starting with data annotation.
The auto-labeling platform grows into the workhorse, processing large volumes of multimodal data for CV and autonomous-driving clients.
Stardust closes roughly $6.88M to expand its platforms and data services.
At the World Artificial Intelligence Conference in Shanghai, Stardust unveils the MorningStar SaaS beta and a "Navigator" plan offering free educational access to researchers.
It provides data annotation, data management and custom dataset services to AI companies - turning raw multimodal data into high-quality training datasets, with deep expertise in autonomous driving and large language models.
Rosetta (auto-labeling/annotation), MorningStar (data management engine), COSMO (dataset pyramid for LLMs), plus custom datasets and an LLM instruction-following evaluation service.
Global AI and automotive enterprises including SAIC Motor, BYD, Geely, Bosch, ZF, Xiaomi and JD.com, among others.
Founded in 2017, associated with San Francisco, California, with operations spanning the US and China.
A Series A of roughly $6.88M (reported up to about $7.2M total), led by Meridian Capital Asia and closed around December 2022.
Watch: search "Stardust AI MorningStar demo" and "Stardust AI WAIC 2024" on YouTube for the latest product walkthroughs and conference interviews. No official channel is publicly listed at time of writing.