Building AI that works once is relatively easy. Building AI you can trust every time is a discipline. That’s the central lesson CNH – one of the leading manufacturers of agricultural and construction equipment – learned as its machine learning team grew into an engineering organization operating at the sharp edge of safety-critical systems.
The insight didn’t come from chasing larger models or smarter algorithms but from fixing something far less glamorous: the data pipelines, tracking infrastructure, and reproducibility practices that determine whether anyone can ever trust a promising result again.
Reproducibility is what draws that line. And in an industry where AI decisions carry real consequences, that line is the difference between a model that’s impressive and one that’s genuinely deployable.
In a recent presentation at the AI-Ready Data Summit, Rahul Ramakrishnan, AI engineering manager at CNH, shared a practical, end-to-end AI workflow that moves from the edge to the cloud and back to the edge, and walked through the process CNH uses to bring autonomous tractors to customers.
Supporting AI Use Cases at CNH
CNH uses AI across a wide range of applications, from customer-facing support chatbots to fully autonomous Level 5 systems.
One example is an autonomous tillage tractor that can perform tillage across large farms end-to-end. Safety is built into the system so that when it detects an obstacle, it can come to a safe stop and, in some cases, navigate around it. This is one of the key applications the team supports, particularly by deploying AI on the edge for safety-critical functions.
To make AI work in practice, CNH builds complete systems around its models. That requires a platform that connects data collection, data processing, training, and deployment.
CNH uses an in-house compute platform that supports both AI deployment and data collection. The platform integrates data from cameras, as well as optional LiDAR and radar, to create a complete sensing system that enables safe autonomous operation.
Why does CNH build a platform rather than a one-off feature? The same AI foundation connecting collection, processing, training, and deployment can scale across many applications. The platform can support multiple autonomous systems, from retrofit kits for existing tractors to entirely new robotic platforms.
CNH treats AI as an ecosystem, not a standalone feature.
From the Edge to the Cloud and Back
Everything starts at the edge, where sensors capture everything happening in the field. Data such as GPS, LiDAR, and images is sent to the cloud, where data is prepared and made ready for AI.
Because the real world is noisy and unpredictable, the system has to be designed to operate in messy conditions. For autonomous navigation, especially when no one is in the cab, the system must effectively replace human vision with a safe machine-perception system.
AI combines camera inputs with other modalities, such as LiDAR, to identify targets, assess the environment, and decide whether to stop or navigate around an obstacle. That is the high-level view of the safety system, even though many smaller components work behind the scenes.
Why Data Collection Matters So Much
One of the most critical parts of any AI workflow is data collection. CNH works with roughly 50 terabytes of data, so it has to be very intentional about how quickly data is collected, which machines collect it, and the overall quality of that data.
A major shift in thinking came from realizing that more data is not always better. Early on, the mindset was that storage and bandwidth were cheap enough to upload everything because it might become useful later. Over time, the team learned that quality matters far more than raw volume.
In the team’s experience, 500,000 high-quality, diverse data points can be far more valuable than 5 million redundant ones. True scalability comes from understanding the quality and diversity of the data and optimizing that quality from collection through deployment, not just in post-processing. The team’s view today is that better data, more than better models, has been the most useful success factor in practice.
This is why CNH targets the specific kinds of data it wants to collect. One method is to vectorize each image along its edges and compare its vector with previously collected data. If a new image is very similar to what is already in the dataset, the system may discard it. If it is highly unique, the system saves it.
The team can also search for particular kinds of data by comparing vectors using cosine similarity. For example, if the tractor fails to recognize a person wearing a fluorescent hoodie, the team can vectorize the example, store the vector in the cloud, and distribute it to field machines. Each tractor can then look for similar images and save only unique examples that match the pattern. This allows CNH to intentionally collect edge cases that are causing model failures and fix them more efficiently.
Data is collected on the device and periodically pushed to the cloud. One of the strongest lessons from the workflow is to obsess over data quality. A high-quality collection saves substantial time and money later by reducing cleanup and labeling costs.
Data Wrangling and Discoverability
Once the system collects data, it has to make that data discoverable and usable. CNH organizes data by time, place, and operation, and also vectorizes images so it can index them by feature similarity. That transforms raw storage into a usable data asset.
Starting with Data Vectorization
Vectorization helps the team find examples related to edge cases, measure dataset diversity, and understand the overall makeup of the data. If teams cannot find the right data quickly, model improvement slows down because they spend too much time manually searching through the workflow.
Labeling: The Most Expensive Step
The next major stage is labeling, which is one of the most expensive and time-consuming parts of the process. As the operation matured, CNH evolved from fully manual labeling toward an AI-assisted workflow.
Once data lands in the cloud, AI models auto-annotate much of it. Larger foundation models that are too heavy to run on the edge can be run in the cloud and adapted for the task at hand. These models can label data with roughly 90-95% accuracy, while human labelers adjust the remaining errors or uncertain cases.
This significantly reduces labeling costs, which in many organizations can reach tens of thousands of dollars per month. Tools like SAM for segmentation or DINO for object detection are examples of existing models that can help automate labeling and save time and money. After human review and correction, the final labeled data is stored back in the cloud.
Training Is Not the Hardest Part
In CNH’s view, training is actually one of the easier parts of the pipeline. Training routines are often fixed, and much of the work comes down to hyperparameter tuning or testing different model-dataset combinations.
By contrast, labeling has a much more direct impact on model behavior. One key recommendation is to be extremely detailed when designing labeling schemes. For example, if the team doesn’t initially tag images by conditions such as sunny, rainy, or cloudy weather, it becomes much harder and more expensive later to evaluate model performance across those conditions.
Labeling projects should therefore include as many useful tags as possible from the start, because labels shape how models behave and how performance can be analyzed.
Tracking and Reproducibility with lakeFS
During training, CNH generates hundreds, or even thousands, of candidate models across different datasets, configurations, and approaches. While many people assume training is where the “AI magic” happens, the team sees data that feeds those models as the harder problem – how to understand exactly which data was used where and how to ensure reproducibility. This is where lakeFS becomes an important part of CNH’s architecture.
CNH uses lakeFS to version data the same way software teams version code. The team branches, commits, compares, and promotes datasets so that every experiment is tied to a specific data snapshot. This allows the team to recreate experiments, compare results, and understand exactly what changed.
Instead of guessing why a model improved or regressed, the team can trace outcomes back to specific data versions.
During raw data ingestion and labeling, separate lakeFS branches are created for different experiments, hyperparameters, or dataset variants. Since lakeFS provides zero-copy isolation, data can be transformed and then used for training on a specific commit. This creates a full lineage trail for each experiment.
In combination with tools like MLflow for MLOps tracking, lakeFS helps CNH track the lineage of the data itself.
Why Reproducibility Is Critical
In a safety-critical system, reproducibility is essential. The team must be able to reproduce a successful experiment, make controlled changes, and understand why those changes affected performance.
Without reproducibility, any model’s success is left to chance. In a system where failures could affect life or property, this is not an acceptable foundation. Reproducibility is also a compliance requirement for autonomy standards. If the team cannot reproduce a result, it cannot fully trust the result or move forward with deployment.
In practice, training may generate 800 or more candidate models. The team then selects the most generalizable candidates with the strongest metrics and subjects them to behavioral tests. These tests may use historical field logs and tens of thousands of images containing critical targets that must be detected reliably.
Only after a model performs well in those evaluations does the team further verify its performance. The key question after training is not simply whether a model worked, but why it worked. If the team cannot answer that, it cannot improve the model intelligently or trust it enough for production use.
AI as a Continuous Loop
Once a model is deployed, it returns to the edge, where it begins generating new data, scenarios, and insights. Those feed into the next iteration of model improvement. CNH therefore treats AI as a loop, not a linear pipeline.
Model performance can be improved in several ways, including scaling model size, improving augmentations, or adding compute.
But from the team’s experience over years of building real-world autonomous tractor systems, the biggest lever is data. Better data produces better models, better outcomes, and better iteration cycles. In this workflow, better data matters more than bigger models.
Main Takeaway
The main takeaway from the talk was that success doesn’t come from smarter or larger models alone. It comes from better data systems and strong tracking and reproducibility.
CNH’s focus is on moving toward reliable, safe models built on reproducible processes. That shift reflects the team’s evolution from initial experiments to the disciplined practices required in a mature company operating safety-critical AI systems. In CNH’s experience, reproducibility is what turns promising results into trustworthy success.
Want the full story? Watch Rahul’s session on demand from the AI-Ready Data Summit.
