Executive Summary

For the past few years, attention in AI has fixed on the model. Bigger models, longer context, more elaborate architectures were the axis of competition. But as that race matured, the variable that decides performance quietly moved. Given the same model, what usually changes the outcome now is the data. The bottleneck has passed from model to data.

Andrew Ng is the one who put a name on that shift. He defined Data-Centric AI as the discipline of systematically engineering the data used to build an AI system. Instead of tweaking the model, hold the model fixed and iterate on the data. There is already a clinical logic inside that idea: check the state, intervene, check the state again. Diagnose, prescribe, re-diagnose.

So when people talk about "fixing" data, they reach naturally for medical vocabulary. They say the data is sick; they say they diagnose it. The question is how seriously, and how far, that metaphor actually runs. This report asks two things. How is the view of data-as-medicine scattered across industry and academia right now? And on that map, where exactly does Pebblous stand?

Data quality has borrowed from the language of hygiene and medicine since its origins. Talk of "cleaning" and "scrubbing" the "dirty data" was already idiom back in the data-warehouse days of the 1990s. In CRM and security databases, the field went a step further and began treating data like a living organism, with terms like "data hygiene," "data health," and "data decay."

Most of this vocabulary, though, is idiom rather than method. Even DAMA-DMBOK, the field's standard reference, defines data cleansing operationally in terms of accuracy, completeness, and consistency, not medicine. The metaphor sits on the labels; it was never built into the methodology. The table below sorts the medical and hygiene vocabulary that has seeped into data by character.

Of these, "quarantine" and "triage" have settled past metaphor into real engineering patterns. Routing anomalous records into a separate table is standard pipeline design by now, and the "sort it, fix it, let it flow again" procedure is widely used in streaming-data recovery. In short, the vocabulary of medicine and epidemiology has already been grafted onto data practice. But it lives as scattered idioms or isolated patterns; nothing yet ties them into a single, integrated clinical methodology.

The side that industrialized the medical metaphor most is the data observability business. Monte Carlo, which opened the field, coined "data downtime" by analogy to a server going down, organized data state into the "5 pillars of data health," and put "detect, resolve, prevent" on the marquee. Other vendors then adopted "data health" as product language.

Yet of the nine vendors surveyed, seven stop at diagnosis and monitoring. What they call "resolve" or "remediation" is almost always fixing an incident, a pipeline, or a job, not treating the data values themselves. Trace the real lineage of the metaphor and it lands closer to software observability (DevOps) than to medicine. "Downtime" and "uptime" are the language of servers; only "health" among them was borrowed in a quasi-medical sense. The table below sets each vendor's metaphor against its actual scope.

Source: a full review of the official documentation of nine observability vendors (Monte Carlo, Great Expectations, Bigeye, Anomalo, Soda, Databand, Datafold, Metaplane, Acceldata). The table shows a representative six.

There are two exceptions. Soda puts treatment out front — "your data has a problem, now it fixes itself" — and Databand attaches automated action. But some of Soda's language is still at the announcement stage, and what Databand acts on is jobs and pipelines, not data values. The two moves are signals that the gap is narrowing, yet both have only reached the entrance. Screening and emergency sorting are industrialized; the operating room and rehab, treating the data back to health and re-diagnosing it, remain empty.

The place on the map closest to Pebblous in philosophy is Data-Centric AI (DCAI). By Ng's definition, it is the discipline of holding the model fixed and systematically iterating on the data. The circular thinking is already inside it: look at the state, intervene, look again. A loop that runs from diagnosis to improvement and back to reassessment.

The movement began with a 2021 talk, became the Data-Centric AI workshop at NeurIPS that year, and was then institutionalized as DataPerf, a benchmark that competes on datasets. DataPerf's premise compresses into one line: iterate on datasets, instead of just architectures. Zha et al.'s 2023 survey organized the movement into a "cyclical, model-in-the-loop" pipeline spanning the development and maintenance of training and inference data.

Read the originals, though, and one thing is clear. Data-Centric AI expresses data improvement as "systematic engineering" and "iterating on datasets"; it does not use the medical vocabulary of diagnosis, treatment, clinic, or patient. Medicine appears only as an example application domain. The structure of the loop is essentially the same shape as Pebblous, but the metaphor layer is left blank.

Academia has a rich supply of methods for diagnosing data. Here too, though, the medical metaphor usually rides on the title, while the body of the work speaks the neutral language of engineering and statistics. Consider four representative threads.

Dataset Cartography (Swayamdipta et al., 2020) says so in the title itself — "Mapping and Diagnosing." It uses training dynamics to separate easy, ambiguous, and hard regions and draw the terrain of a dataset. Confident Learning (Northcutt et al., 2021) and its implementation, Cleanlab, find label errors statistically and became an industry standard. TRIAGE (Seedat et al., 2023) borrows emergency medicine's sorting for its name, but the body focuses on characterizing and auditing training data for regression. IBM's Data Quality Toolkit (Gupta et al., 2021) comes closest to a closed loop with its "detect, explain, remediate" procedure.

Three things stand out. First, essentially no paper uses "diagnose, symptom, treat, health" consistently as its working verbs. The brand seat for the clinical metaphor is empty in academia too. Second, no paper closes the clinical loop from diagnosis to treatment and back to re-diagnosis. The IBM toolkit comes closest but sets no explicit "re-diagnosis" step. Third, some pieces, like label-error detection, are already mature enough to be an industry standard. The materials for diagnosis are plentiful; what is missing is the seat where they are stitched into one clinical procedure.

Once the diagnosis is done, a prescription has to follow for the clinical work to be complete. Treating data splits into two broad directions: a diet that trims the excess, and a bulk-up that fills what is missing. Each direction is backed by a mature, independent line of research.

6.1Diet — Treatment by Subtraction

The evidence that piling on more data is not always best is clear. Sorscher et al. (2022) showed that pruning data by a good metric can push performance past power-law scaling toward exponential gains. Lee et al. (2021) found that simply removing duplicates from training data makes language models better and sharply reduces memorization. Strategically trimming overlapping, common data lowers training time and cost, and the density of what remains actually raises quality.

6.2Bulk-Up — Treatment by Addition

On the other side is synthesis. Liu et al.'s 2024 survey lays out synthetic data as an effective tool for scarcity, privacy, and cost, while attaching the condition that factuality and fidelity must be verified. Fill without verifying and synthesis makes the data sicker, not healthier. That is why bulk-up necessarily drags re-diagnosis along with it: unless you image again after filling, you cannot know what you filled in.

The framework that ties these two prescriptions under one higher concept is the Zha et al. 2023 survey seen earlier. By organizing data development as a cyclical, model-in-the-loop pipeline, it lets us view diagnosis, treatment, and maintenance as a single flow. Since 2024, autonomous loops such as the "data engine" and "data flywheel," which mine failure cases to retrain, have also appeared. But these mostly stay in a narrow loop of relabeling or cleaning, some distance from a complete cycle that holds diagnosis, diet, bulk-up, and re-diagnosis in one pipeline.

The landscape drawn so far had its materials scattered around: observability's "health," quarantine and triage, pruning and synthesis, Data-Centric AI's iteration logic, and Van den Broeck's diagnose-then-treat procedure. What Pebblous did was join these pieces into a single closed clinical loop and make it measurable.

7.1DataClinic — The Instrumented Diagnostic Platform

DataClinic is Pebblous's AI data-quality management platform, automating the full path of diagnosing, improving, and certifying data-quality issues. The core is turning quality from a "feeling" into a "number." Data Imaging technology moves raw data into high-dimensional embeddings, translating quality characteristics into geometric signals of density, distance, and distribution. That is why Pebblous likens the process to ultrasound or MRI: it images inside a dataset the way you look inside a body, locating the defects that correspond to lesions.

That measurement needs a reference. Pebblous takes the "what" defined by ISO/IEC 5259, the international standard for data quality, and moves it into an actually measurable "how." If the standard is a legal code, DataClinic is the instrument that enforces it. This measurement technology was designated a 2025 Innovative Product by Korea's Public Procurement Service and is protected by a U.S. patent (US 12,481,720 B2).

7.2Two Prescriptions and the Re-Diagnosis Cycle

The diagnosis leads straight into a prescription. For high-density regions where similarity and duplication run excessive, it prescribes a data diet, strategically trimming unnecessary data to lower training time and storage cost. For low-density "gap" regions short on representativeness and diversity, it prescribes a data bulk-up, identifying missing edge-case regions and filling them with targeted synthetic data. And once the prescription is done, it images again. "Diagnose with DataClinic, generate precisely with PebbloSim, then re-diagnose with DataClinic": that cycle is the structure Pebblous calls the Data Flywheel.

Note: Pebblous's own canon does not name "diagnose → diet → bulk-up → re-diagnose" as a single one-line definition; it describes two branching prescriptions based on the diagnosis plus a separate re-diagnosis cycle. This diagram combines the two into one loop.

7.3Data Greenhouse — From Clinic to Autonomous Operation

If DataClinic is the diagnostic tool, Data Greenhouse is the next step that extends that tool into an autonomous operating system. Here the metaphor moves from medicine to agriculture. It is a declaration to treat data not as prey you hunt down once, but as a crop that grows on its own given the right conditions and grows more fertile the more it is tended. Hunting takes from nature; cultivation grows with it.

Data Greenhouse aims at a data operating system that autonomously observes, judges, acts, and proves. It does not stop at a one-off diagnosis but sustains the cycle of diagnosis and improvement. The execution engine that makes this possible is PebbloSim, which combines physics simulation with generative AI in a digital-twin virtual environment to produce high-quality synthetic data free of physical hallucination. And it records the whole compliance path, from ISO 5259 diagnosis through the EU AI Act and ISO 42001, as "Operational Evidence." The table below lays out the relationship between the two concepts.

It is honest to make one thing clear at this point. Pebblous did not invent data medicine. As we saw, the diagnose-then-treat clinical procedure was already in Van den Broeck's paper 20 years ago; the individual techniques of diet and bulk-up, and the philosophy of iterative improvement, each stand on established literature. "We were first" does not hold on this map.

So it is more accurate to describe Pebblous's position as "integration" rather than "invention." It is the first to tie the scattered pieces into three things. First, a clinical loop that closes from diagnosis through prescription (diet · bulk-up) and back to re-diagnosis. Second, a direction where that loop is run autonomously by Agentic AI rather than by hand (from DataClinic to Data Greenhouse). Third, a discipline that measures the whole process against a standard, ISO/IEC 5259. Each piece rests on prior literature; combining these three into one, integration, autonomy, and standardized measurement, is the new recombination. Honestly acknowledging the precedent is exactly what makes this distinctiveness defensible.

The models are mostly out. The remaining bottleneck is data, and the way we handle data is moving past cleaning toward something closer to clinical care. The seat that joins diagnose, prescribe, and diagnose-again into one and runs it autonomously, the seat that many pieces of the landscape point toward yet no one has fully filled, is where Pebblous means to stand.

Academic Papers

• 1.Van den Broeck, J., Argeseanu Cunningham, S., Eeckels, R., & Herbst, K. (2005). Data Cleaning: Detecting, Diagnosing, and Editing Data Abnormalities. PLoS Medicine, 2(10):e267.
• 2.Zha, D., et al. (2023). Data-centric Artificial Intelligence: A Survey. arXiv: 2303.10158.
• 3.Mazumder, M., et al. (2022). DataPerf: Benchmarks for Data-Centric AI Development. arXiv: 2207.10062.
• 4.Swayamdipta, S., et al. (2020). Dataset Cartography: Mapping and Diagnosing Datasets with Training Dynamics. EMNLP. arXiv: 2009.10795.
• 5.Northcutt, C., Jiang, L., & Chuang, I. (2021). Confident Learning: Estimating Uncertainty in Dataset Labels. JAIR. arXiv: 1911.00068.
• 6.Seedat, N., et al. (2023). TRIAGE: Characterizing and Auditing Training Data for Improved Regression. NeurIPS. arXiv: 2310.18970.
• 7.Gupta, N., et al. (2021). Data Quality Toolkit: Automatic Assessment of Data Quality and Remediation for Machine Learning Datasets. arXiv: 2108.05935.
• 8.Sorscher, B., et al. (2022). Beyond Neural Scaling Laws: Beating Power Law Scaling via Data Pruning. NeurIPS. arXiv: 2206.14486.
• 9.Lee, K., et al. (2021). Deduplicating Training Data Makes Language Models Better. arXiv: 2107.06499.
• 10.Liu, R., et al. (2024). Best Practices and Lessons Learned on Synthetic Data for Language Models. arXiv: 2404.07503.
• 11.Floridi, L., & Cowls, J. (2019). A Unified Framework of Five Principles for AI in Society. Harvard Data Science Review.

Industry Documents & Standards

• 12.Ng, A. (2021). A Chat with Andrew on MLOps: From Model-centric to Data-centric AI. DeepLearning.AI. (definition of Data-Centric AI)
• 13.DAMA International. DAMA-DMBOK: Data Management Body of Knowledge. (dimension-based definition of data cleansing)
• 14.Moses, B., et al. Monte Carlo. What Is Data Downtime? The 5 Pillars of Data Health. montecarlodata.com.
• 15.Observability vendor documentation (full review): Great Expectations, Bigeye, Anomalo, Soda, Databand (IBM), Datafold, Metaplane, Acceldata.
• 16.ISO/IEC 5259 (2024–). Artificial intelligence — Data quality for analytics and machine learning (ML).
• 17.Pebblous DataClinic & Data Greenhouse in-house documentation. blog.pebblous.ai. (Patent US 12,481,720 B2; 2025 PPS Innovative Product)

Note: recent work since 2024 on autonomous data loops (data engines and flywheels) is described in the body only as a trend, without asserting any specific paper. Andrew Ng's directly quoted phrasing is a paraphrase that would need to be re-checked against the slide deck and video.