8M+ annotations advancing AI readiness at scale

HCLTech built an AI Data Factory for a global leader in AI computing. The initiative achieved a 95% quality score and a 47% efficiency improvement, accelerating LLM training

5 min read

Published: 9 November 2026

Last updated: 9 November 2026

Overview

A global leader in AI computing and accelerated computing platforms partnered with HCLTech to build and operate a production-grade AI Data Factory. The engagement integrated multimodal data operations, LLM training, Physical AI and robotics data, model alignment and safety validation. Together, these capabilities transformed raw data and real-world signals into high-fidelity training data.

The Challenge

The growing scale of AI drove the need for enterprise-scale training data operations

As AI moves beyond text and images into robotics, autonomous systems and other Physical AI applications, training data must do more than provide volume. It needs to be accurate, contextually rich, human-validated and safe.

For the client, rapidly expanding AI workloads created a need to industrialize data operations across multiple model types and real-world use cases. It needed an enterprise-scale data operation capable of continuously converting diverse data into trusted, production-ready training signals.

The operation needed to bring together:

Multimodal data ingestion and structuring at scale
Human reasoning, preferences and domain expertise
Real-world datasets for Physical AI and VLA models
Continuous quality measurement and improvement
Model evaluation, red-teaming and safety validation

The Objective

Scale trusted AI data operations through industrialized delivery, embedded quality and robust governance

The engagement focused on six priorities:

Production-grade AI Data Factory: Build and operate an end-to-end data lifecycle, from raw ingestion to production-ready model training datasets, at enterprise scale.
LLM training data excellence: Deliver high-quality text, multimodal and domain-specific training data rated for content safety, factual accuracy, coherence and helpfulness.
Physical AI and VLA model readiness: Enable next-generation Vision-Language-Action models and robotics systems with real-world, human-validated egocentric and teleoperation datasets.
Human-in-the-loop model alignment: Bring human reasoning, preference and domain expertise into model learning through RLHF, preference scoring and instruction tuning workflows.
Embedded quality management: Operate a continuous quality audit program using a five-star methodology and quality pass-rate tracking to drive training interventions and raise annotation standards.
Safety, governance and trust: Establish structured red-teaming, adversarial testing and guardrail validation frameworks to support Responsible AI deployment at scale.

The Solution

An enterprise-scale AI Data Operations capability powered by a governed AI Data Factory

Deployment and scale: HCLTech deployed 500+ specialists across LLM, multimodal, Physical AI, robotics and GenAI workflows, operating a scalable AI Data Factory from capability centers in the Philippines and India.
LLM training and content operations: Human reviewers rate model outputs across content safety, factual accuracy, coherence and helpfulness, while specialists produce structured prompts and captions across English, Korean, Japanese and Vietnamese.
Physical AI, robotics and VLA models: Produced egocentric first-person video data, teleoperation datasets, simulation environments and autonomous vehicle decision-reasoning traces to support Vision-Language-Action (VLA) models and robotics systems.
Multimodal data ingestion and structuring: Created structured extraction pipelines for complex documents using bounding-box recognition, intelligent noise filtering and dense video transcriptions.
Model alignment and evaluation: Applied multi-axis preference scoring to model responses and evaluated synthetic RAG queries, retrieved documents and generated responses for relevance, completeness and factual grounding, helping reduce hallucinations.
Safety, governance and red-teaming: Deployed adversarial testing workflows to identify jailbreaks, policy violations and unsafe outputs before production release.
Quality management framework: Operated a calibration-driven QA model featuring a five-star audit methodology, weekly review loops and continuous quality pass rate tracking.
Domain specialist talent model: Recruited professionals with specialized subject-matter expertise across gaming, creative writing, computer science, finance and medical sciences to bring domain knowledge and human reasoning into training data.

The Impact

A scalable data operations capability that moves AI from training to readiness

The AI Data Factory delivered measurable gains in quality and efficiency while establishing the scale needed to support the client's expanding AI development efforts.

95% quality score achieved across annotation and data-labeling workflows through the five-star audit framework.
47% process efficiency improvement delivered through continuous calibration, weekly QA loops and targeted training interventions.
8M+ annotations produced annually across LLM, Physical AI, multimodal and domain-specific workstreams.
500+ specialists deployed across capability centers in the Philippines and India.
Accelerated AI readiness: Cleaner, structured datasets built using the pyramid principle and domain expertise accelerated the path from raw data to production-ready AI models.
Stronger AI safety: Structured red-team validation and RAG and RLHF evaluation strengthened safety and governance across AI workflows, helping identify potential risks ahead of production deployment.

The engagement demonstrated how an enterprise-scale AI Data Factory can turn data operations into a strategic capability for advancing trustworthy AI.