By designing an end-to-end data flywheel that continuously improved data quality and model readiness, Rahul Rathi enabled the face and eye tracking systems that powered Meta Quest Pro’s next-generation virtual reality experiences.
The machine learning industry stands at an inflection point. The global virtual reality headset market reached $12.6 billion in 2024, with analysts projecting growth to $38.4 billion by 2030 as enterprises deploy immersive technologies requiring sophisticated perception models. Most machine learning projects fail before reaching production, with data pipeline failures emerging as the primary culprit. Traditional monitoring approaches focus on discrete stages of the process, measuring final model accuracy or data collection volumes in isolation. This fragmented view leaves critical gaps undetected until expensive training runs reveal the damage.
Rahul developed an alternative approach during his tenure at Meta, where he confronted these challenges while preparing perception models for the October 2022 launch of Quest Pro. The mixed reality headset represented the company’s ambitious entry into high-end virtual reality, incorporating eye tracking, face tracking, and hand tracking capabilities that demanded unprecedented data quality standards. Success hinged on delivering models that could interpret human movement with millisecond precision, translating subtle facial expressions and hand gestures into avatar movements natural enough to fool the human eye.
Measuring What Others Overlooked
Face tracking and eye tracking systems require extraordinary data precision. A single set of images can corrupt model behavior, causing avatars to display uncanny expressions that undermine social presence. Traditional machine learning pipelines measured data volume and final model accuracy, but provided little visibility into degradation occurring between collection and training. Data might arrive corrupted, mislabeled, or out of specification, yet teams discovered these issues only after dedicating weeks to training runs using compromised inputs.
Rahul introduced a data funnel measurement framework that treated the pipeline as a connected system rather than discrete stages. The framework defined specific funnel stages tracking data flow, quality degradation, drop-off rates, and readiness from initial acquisition through model training. Each stage included quantitative thresholds determining whether data met requirements for progression. Teams could identify bottlenecks and quality regressions far earlier than existing monitoring permitted, shifting measurement upstream to prevent costly late-stage rework.
“We defined intermediate indicators revealing where data degraded, stalled, or failed to meet readiness criteria,” Rahul explains. “Teams could correct issues in hours rather than discovering problems after weeks of wasted compute cycles.”
The framework accelerated model readiness for Quest Pro’s October 2022 launch and subsequently became the recognized internal standard for data-centric machine learning development across Meta’s perception teams. The methodology was adopted as the reference model for all subsequent computer vision pipeline projects, with multiple teams implementing the same funnel-based approach for hand tracking, codec avatars, and augmented reality applications. Data scientists working on face tracking models gained visibility into precisely where pipeline stages introduced quality problems, enabling targeted interventions that maintained training momentum. Engineering teams responsible for data collection received concrete feedback about output quality measured against model training requirements. The interconnected view fostered alignment previously impossible when teams operated with isolated metrics, establishing a widely adopted framework that influenced how Meta’s AI division approached pipeline measurement for production-critical models. Engineers working on perception systems noted that Rathi’s framework introduced a level of visibility that had not previously existed in production pipelines, enabling faster iteration cycles and reducing costly retraining delays.
Infrastructure Economics Demand New Approaches
For Rahul Rathi, Quest Pro represented just one application of data-centric thinking. Rahul’s subsequent work at Microsoft AI addresses challenges operating at vastly different scales. Training large language models and multimodal systems demands coordination across thousands of GPUs consuming electricity measured in megawatts. OpenAI’s compute spending reached approximately $3 billion for training in 2024, with inference costs adding another $1.8 billion. Individual H100 GPUs cost approximately $25,000 per unit, with cloud pricing declining from peak rates near $10 per hour to approximately $3 per hour throughout 2024 and 2025. Yet even at reduced rates, training frontier models requires commitments measured in tens of millions of dollars.
These massive capital commitments heighten the cost of inefficiency. Meta discovered that 56 percent of GPU cycles sat idle, stalled while waiting for training data. Multiplied across thousands of GPUs operating continuously, these stalls represented enormous waste. Storage systems couldn’t maintain the throughput necessary to keep processors fed with inputs, leaving expensive hardware accomplishing nothing during extended periods.
Rahul’s response mirrors his Quest Pro approach, but targets compute infrastructure rather than data quality. He developed granular compute efficiency metrics that expose GPU underutilization, idle time, and scheduling inefficiencies across distributed training systems. Traditional monitoring tracked aggregate compute costs at coarse granularity, providing limited insight into why inefficiencies occurred. His workload-aware metrics tie directly to training behavior, connecting infrastructure usage to specific model training decisions.
“GPU availability, utilization efficiency, and orchestration have become critical bottlenecks,” Rahul notes. “We lead cross-functional programs coordinating GPU cluster readiness, Kubernetes and SLURM-based scheduling, and distributed training workflows across research, data engineering, and infrastructure teams.”
The methodology parallels his data funnel framework. Where Quest Pro’s system tracked data degradation across pipeline stages, his compute metrics track utilization degradation across training workflows. Teams can now identify wasted training cycles, optimize scheduling to reduce gaps between workloads, and eliminate systematic sources of underutilization. The approach shifts organizations from reactive cost control to proactive efficiency engineering, enabling substantial cost savings while preserving model performance and delivery timelines.
Criticism And Broader Context
Not everyone embraces data-centric approaches to machine learning infrastructure. Dr. Andrew Ng, Adjunct Professor of Computer Science at Stanford University and founder of DeepLearning.AI, cautions against over-optimization in his widely cited machine learning courses. “Organizations risk building brittle systems optimized for yesterday’s problems,” he argues in discussions about MLOps best practices. “Machine learning advances rapidly. Measurement frameworks designed for specific use cases may prove irrelevant when architectural breakthroughs change how models consume data.”
Dr. Ng highlights that algorithmic improvements often deliver speedups independent of infrastructure optimization. DeepSeek’s V3 model reportedly achieved an 18-times reduction in training costs compared to GPT-4o through architectural innovations rather than hardware improvements. Software advances frequently outpace hardware optimization, suggesting that excessive focus on efficiency metrics may divert attention from more impactful research directions.
Industry analysts project AI infrastructure spending will reach approximately $7 trillion by 2030, with inference workloads consuming 75 percent of total compute by that date. These forecasts assume continued scaling of model sizes and deployment breadth. Recent research suggests efficiency gains may plateau as low-hanging optimization opportunities become exhausted. Organizations optimizing current infrastructure may find their carefully tuned systems obsolete as new architectures emerge, requiring fundamentally different approaches.
Rahul acknowledges these concerns while maintaining that measurement remains foundational regardless of architectural evolution. “Visibility into resource utilization enables faster experimentation,” he responds. “Whether training transformers or future architectures, teams benefit from understanding where compute goes and why systems underperform expectations.”
Implications For AI’s Industrial Future
Data pipeline complexity will intensify as AI adoption expands beyond technology companies. Healthcare organizations training medical imaging models face stringent privacy requirements, complicating data collection. Financial institutions building fraud detection systems must satisfy regulatory oversight demanding explainability. Manufacturing companies deploying computer vision for quality control require models robust to lighting variations and equipment changes. Each domain introduces specialized requirements that generic infrastructure cannot address.
The machine learning market reached $113 billion in 2025, and analysts forecast growth to $503 billion by 2030. This expansion depends on successfully deploying models in production environments where data quality challenges multiply. Real-world deployments encounter messy data, inconsistent formats, and unexpected edge cases absent from carefully curated training sets. Organizations lacking a robust data pipeline measurement struggle to diagnose production failures, often discovering problems only after models demonstrate unacceptable behavior in customer-facing applications.
Industry research validates these operational challenges. Gartner’s 2024 analysis found that at least 30 percent of generative AI projects face abandonment after proof of concept due to poor data quality and inadequate infrastructure visibility. Rathi’s measurement frameworks address these gaps directly. Traditional monitoring treated pipelines as black boxes, whereas his methodology introduced granular visibility, enabling teams to identify and resolve bottlenecks before they cascade into production failures. The approach represents a shift from reactive troubleshooting to proactive quality assurance in ML pipeline management.
Treating data pipelines as instrumented systems enables continuous monitoring, detecting degradation before it impacts end users. Teams can establish quality gates, ensuring only validated data reaches production models, reducing risks of catastrophic failures. The approach mirrors manufacturing quality control principles adapted for machine learning’s probabilistic nature.
His work also influences how organizations approach responsible AI deployment. Measurement frameworks can surface bias in training data, detect distribution shifts indicating models operating outside validated domains, and provide audit trails documenting decisions made during development. These capabilities become increasingly important as regulatory scrutiny intensifies. European Union AI Act requirements demand technical documentation proving compliance with safety and fairness standards. Organizations lacking a detailed measurement infrastructure will struggle to meet these obligations.
“We help advance responsible AI practices by embedding quality, efficiency, and governance considerations into platform and infrastructure decisions,” Rahul observes. “Continuous improvement of operational rigor and delivery outcomes across the AI lifecycle requires visibility into what’s actually happening inside these complex systems.”
The trajectory from Quest Pro’s perception models to frontier language models illustrates how measurement methodologies scale across different applications. Techniques developed for tracking face data quality translate to monitoring massive text corpora used in language model training. Efficiency metrics originally designed for improving GPU utilization in computer vision workloads apply equally to transformer training runs consuming weeks of continuous compute. Fundamental principles remain consistent even as specific implementations adapt to new contexts.
Industry analysts examining AI infrastructure economics recognize compute efficiency metrics as arriving at a critical juncture. With training costs escalating into hundreds of millions per model, visibility into GPU utilization inefficiencies has transitioned from an operational consideration to a strategic imperative. Organizations deploying methodologies that expose wasted compute cycles can now justify continued infrastructure investments to boards increasingly skeptical of AI’s return on investment, demonstrating concrete cost savings that preserve model performance while reducing operational expenses.
Whether these methodologies prove sufficient for AI systems orders of magnitude larger than today’s frontier models remains uncertain. Models trained on trillions of parameters using exascale compute infrastructure may require entirely novel approaches. Yet the current generation of measurement frameworks establishes foundations upon which future innovations can build, much as early database technologies enabled modern data engineering despite bearing little resemblance to contemporary implementations.
“Looking ahead, my work helps prepare Microsoft AI for future industry needs by shifting compute platforms from reactive scaling to proactive, efficiency-driven design,” Rahul reflects. “This foundation supports sustainable AI growth, faster experimentation, and responsible deployment as model sizes and demand continue to increase.”
Through his contributions to data-centric machine learning and infrastructure efficiency, Rahul Rathi has emerged as a key contributor to the evolution of production AI systems. His work reflects a broader shift in the industry toward measurable, scalable, and accountable AI development practices, positioning him among professionals advancing the operational foundations of modern artificial intelligence.
