Unique Research · Ranking / Unique Friend interview · January 8, 2024 · Author (CN): Xu Meng · Guest: Zhang Lei, Founder, Stardust Data (星尘数据)
Inventory type kept as ranking per eligible inventory (Unique Friend / company interview package; not partner welcome).
Stardust Data builds on an AI full-lifecycle data closed-loop system — data production, data management, hard-example discovery, model inference — and develops ML-driven automated labeling tools to inject tech into the labeling industry and empower AI firms’ algorithm iteration and data loops.
Founder Zhang Lei spoke with Unique Research on labeling-industry status, Stardust’s edges, and a data-centric strategy.
Labeling market: large growth room
AV and LLMs have driven explosive labeling demand. Global labeling is projected to grow at ~40% CAGR, with market scale expanding fast. Labeling flows themselves are being remade by the foundation-model paradigm — from manual toward algorithm-driven auto-labeling — lifting efficiency and widening the market further.
Earlier, China’s labeling industry had a low entry bar and was demand-side led: suppliers built tools for large annotator pools, task-oriented, relying on channels and labor — a traditional mode.
GenAI raised requirements: buyers need stage-aware (not only task-aware) support, strong pre-label and assist algorithms — and, harder still, products that host those algorithms and close the data loop into data collaboration. Difficulty and labor cost rose, challenging classical labeling and opening space for algorithm-first auto-label modes.
Zhang founded Stardust in 2017 with three goals: (1) raise labeling efficiency via automation; (2) create more customer value via data-strategy expert services; (3) close data loops so model training works better with less labeling volume — better cost/performance.
Zhang: undergrad Zhejiang University, master’s George Washington University; 10+ years Wall Street and Silicon Valley — built what CN calls the world’s first equity-investment robot in Silicon Valley; risk algorithms from Wall Street work used as stress-test standards by the U.S. Federal Reserve (per CN); expert committee member, National Voice & Image Product Quality Inspection Center; AV expert, China Productivity Promotion Center Association. In 2017 his team shipped China’s first commercialized automated labeling platform.
Customers are growing ~2× per year, mainly LLM and AV. On perception/decision, Stardust serves ~90% of leading AV firms; AGI / large-model logos are rising, with multiple top LLM customers signed.
Auto-assist labeling + data-loop iteration
1. Rosetta — China’s first auto-labeling platform (~70% automation)
Core business: AI full-lifecycle data closed-loop base covering production, management, training, deployment. Two curves: (1) keep raising labeling-platform automation to lift AI R&D efficiency; (2) data-loop systems + LLM data solutions for vertical pain points.
Rosetta — claimed China’s first automated labeling platform — offers hundreds of labeling functions across 100+ mainstream scenes (4D-BEV, image, point cloud, text, speech, collection, etc.) and 100+ assist algorithms for pre-label, label, and QC. Automation rises year by year.
Edge: self-built Autolabeling — platform automation 70%+, ~200 million records processed/year, 99.9% accuracy, 1,000+ customer projects. Tech, industry experience, and iteration capability sit at the front of the pack (per CN). Higher automation lifts annotator throughput and income vs other platforms.
2. Auto-assist + closed-loop iteration for algorithm speed
A full data loop that covers ingest, orchestration, preprocessing, hard-example discovery, send-to-label, acceptance, training-process management, and model-value metrics can greatly accelerate algorithm iteration. Zhang: “Stardust is not solving 100% of labeling — it solves the hard, frontier, large-scale labeling that paying large customers feel most.”
Data and demand complexity keep rising: 2D recognition → 3D → 2D/3D fusion → BEV 4D — more tech paths, finer rules, higher annotator bars.
Industry responses: (1) labor optimization — vocational schools into the system, cheaper labor, tighter labor control; (2) sales/BD for large sticky accounts and stable margin; (3) marketing that may not profit (even lose money) to grow revenue scale then finance the roll. Classical flow: algorithm teams write rules; labeling firms digest; annotators execute — suboptimal in one pass, and annotators stay “doers” who do not help iterate model quality.
Stardust aims at a fourth path: auto-assist labeling + data-loop iteration to truly cut cost and raise efficiency — tech iteration over years (one algorithm version 1–2 months; real effect may take 1–2+ years). Commercializing across that horizon stresses sales, R&D cadence, and company ops — but Zhang believes algorithms can cut labor cost >90% and sharply lift delivery to meet industry demand.
3. Platform/tools authority + data-strategy experts
Data strategy spans frontier tech: domain transfer, spatiotemporal fusion, augmentation, weak supervision; plus data-centric feedback loops — active learning, RL, retrieval, human-in-the-loop, data security, scene data generation, model testing. Experts must know both tech and product.
After serving ~90% of leading AV customers across 1,000+ projects, Stardust has both a leading platform/tool stack and seasoned data-strategy experts. Versus labor-optimization shops that win an order, split it, and farm to third parties, Stardust’s experts decompose algorithm needs and choose the right way to process customer data.
From labeling toward data infrastructure
1. Raise data reuse; cut LLM training cost
Across the data lifecycle, labeling is a small slice. Shipping an algorithm spans demand definition, solution design, collection, labeling, model design, training, metric tests, inference optimization — each touching data design, use, mining, testing. In the LLM era, massive data management and model reproduction are acute pains.
Stardust’s answer targets those pains. MorningStar supports key links of efficient enterprise AI data iteration — avoiding data-debt pile-up, waste on low-value data, and long feedback chains from train to app — closing the ML loop for a focused, high-speed algorithm production environment.
2. MorningStar — AI full-lifecycle data management
A full ML lifecycle data system sets data quality and use efficiency and caps model upside — key to “data debt” and a future AI competitive moat.
Note (CN): “tech debt” from software — long-term maintenance cost of speed-first builds. “Data debt” is the data-side analogue in ML systems. Weak full-lifecycle data tools slow iteration and accumulate debt.
Against data-debt accumulation, low-value data waste, and long train↔app feedback chains, MorningStar is a one-stop AI full-lifecycle data platform: data landing, orchestration, management, retrieval, hard-example discovery, send-to-label, algorithm management, inference management — continuous iteration speed.
Visual full-lifecycle ops for unified data ops and fast iteration; flexible slicing, multimodal and custom retrieval for special scenes; mainstream hard-example strategies (e.g. active learning); feature distribution, viz/mining, metric compute for algorithm engineers; aggregated data assets for safer collaboration and traceable iteration — breaking internal data silos and lifting AI spend efficiency, accuracy, and data-application capability.
Architecture was closed-loop from day one. First: multi-language/multi-format landing with structured stats — many underlying DBs; unlike classic DMS built for humans, Stardust’s AI DataOps serves algorithms. Proprietary lifecycle management records value discovery from landing through inference automation.
Second: strong storage/management from a decade of team data depth; rich managed objects; extensible hard-example strategies (metrics, custom metrics, distribution/label filters, active learning).
After labeling, the loop continues into algorithm iteration: version management, ground-truth datasets, algorithm evaluation — then closed. Customers can semantically retrieve (e.g. occluded cars), see which data is hot/cold/invalid, which to prioritize for the next algorithm pass, send needed slices to labeling, and feed results back — each iteration recorded so teams need not manually log every data state.
3. Evolving toward a data closed-loop company
Scale AI (2016) pioneered global labeling, then expanded from AV into government, e-commerce, robotics, LLMs — tools, platforms, services. It grew management/test/dev tooling and became a data closed-loop + infra company with broader coverage.
Zhang: Scale’s core is productizing and commercializing tech application — PMF first. Few China startups have real productization; many just ship whatever the customer asks — demand movers, not products. Domestic firms need deep PMs who also grasp frontier research, product purpose, and strategy to connect the go-to-market chain.
Zhang reads Scale’s path as correct: early pricing, strategy, and sales; automation layered in at PMF; brand and data-change systems as the company scaled. That is Stardust’s evolution too — toward a data closed-loop company like Scale.
The “data-centric AI” movement and paradigm shift
1. From model-centric to data-centric
In 2021 Andrew Ng argued the past decade’s biggest AI shift was embracing deep learning; the next decade shifts from model-centric to data-centric. AI = Data + Code; better AI ≈ 80% data + 20% model. Faster ML progress means not only better algorithms but systematic ways to improve data — ~80% of work on data prep; focus from big data to high-quality data.
Data-centric AI is a framework for developing, iterating, and maintaining data in AI systems — building effective training data, designing appropriate inference data, maintaining data — with emphasis on efficient high-quality, large-scale datasets. Stardust’s path empirically follows this idea.
Classic model building iterates models on relatively fixed data (benchmarks + many architectures for accuracy) — model-centric. It underweights real-world issues: wrong labels, duplicates, anomalies. High accuracy may only mean good fit, not good deployment.
Model-centric: iterate models, data relatively fixed. Data-centric: focus on data quality/quantity, model relatively fixed. Data-centric has more real-scene potential because data largely sets the capability ceiling. Core: countless deployment scenes rarely yield “big data”; vs tuning models, it is more feasible to raise sample quality/quantity where the model fails.
2. High-quality data as the performance lever
Engineers often obsess over algorithm design and model metrics and underweight data’s share of the system. Most work — collection, cleaning, features, train, deploy, monitor — is data work. Under the LLM boom, high-quality data lifting performance is consensus; managing volume is the engineer’s pain; data’s role in iteration speed keeps rising.
Getting high-quality data is hard — acquisition, labeling, processing — heavy time and labor. Moving “from big data to high-quality data” needs data-strategy design. OpenAI’s ChatGPT is a famous case of clever strategy — RLHF.
Zhang: the real decisive factor for foundation models is data, not architecture. Same architecture, different data → different effect. High-quality data is today’s and the next stage’s win condition. Data-centric iteration drives AI landing; the old three-step of label → iterate model → deploy is already backward and non-competitive.
After years of research, model design is relatively mature — especially post-Transformer. Raising data quantity and quality becomes the key path to stronger AI systems. When models are strong enough, most people may not train; they design appropriate inference data to extract knowledge. Data-centric R&D will keep pushing AI systems forward.
The data-centric AI movement is driving an industry paradigm shift; data strategy will matter more over the next decade — spanning domain transfer, spatiotemporal fusion, augmentation, weak supervision, and feedback loops (active learning, RL, retrieval, human-in-the-loop, security, scene generation, model testing).
3 CN images (incl. MorningStar UI); attach on ingest. Author (CN): Xu Meng; edit: Qiuping.
— END —