What changed in INNODATA INC's 10-K — 2024 vs 2025

We partner with leading technology companies, frontier AI laboratories, and enterprises to help enable AI systems that perform reliably, align with intended objectives, and operate safely in real-world environments. Our mission is to enable the responsible advancement of artificial intelligence by providing the data, evaluation frameworks, and human expertise required to build AI systems that can be trusted at scale.

We believe that AI will increasingly function as a foundational layer of the digital economy - embedded across consumer products, enterprise workflows, and mission-critical systems.

As AI systems grow more capable and autonomous, we believe the quality of training data, the effectiveness of post-training alignment, and the rigor of ongoing evaluation will be decisive factors in determining whether AI systems are adopted, regulated, and scaled responsibly.

Innodata was founded more than 35 years ago on the principle that high-quality, well-structured data is essential to leading information-retrieval systems. In 2016-2017, we began building proprietary AI language models based on then-emerging research and frameworks and integrating them into our data production workflows.

Through this work, we developed and refined techniques for generating, curating, and validating human-created data used to train probabilistic, learning-based AI systems, and recognized that data quality and structure were critical determinants of model performance.

This insight led us to invest in the development of an integrated set of AI lifecycle data solutions, addressing a growing market need for specialized data engineering, evaluation, and refinement capabilities across the full lifecycle of AI systems.

Today, leading AI innovation labs and Big Tech companies (including five of the so-called “Magnificent Seven”) building frontier generative AI models and leading enterprises engage us to provide (i) training and post-training data development; (ii) alignment and preference optimization; (iii) capabilities, alignment, and safety evaluation; and (iv) AI enablement and operationalization, including support for agentic and tool-using systems.

We believe Innodata is differentiated by: (i) our ability to operate across the AI lifecycle in alignment with AI developers’ internal development and deployment pipelines; (ii) our scale of specialized human expertise; (iii) purpose-built platforms and processes that combine automation with rigorous human oversight; (iv) a research-driven approach to measurement, safety, and operational reliability, which is particularly relevant for frontier model developers and enterprises deploying AI in high-stakes environments; and (v) our dual role supporting leading technology companies building advanced AI systems and enterprises deploying those systems in production, which we believe creates a reinforcing feedback loop that strengthens our capabilities across both contexts and differentiates us from competitors focused on only one side of the market.

Market Opportunities AI Training and Post-Training Data Modern AI systems are trained using large volumes of data rather than explicit, rule-based programming. Foundation models - such as large language models (“LLMs”) and multimodal models - learn statistical representations of language, images, code, and other modalities from vast training corpora.

As model architectures have matured, leading developers have increasingly emphasized the importance of training data quality, data provenance, supervised fine-tuning, and post-training alignment techniques.

We believe that as model scale increases, marginal improvements in data quality and post-training signals can have an outsized impact on performance, reliability, and usability - often exceeding the impact of further parameter scaling alone. 4 Table of Contents Organizations developing AI systems therefore require partners that can design, execute, and continuously refine data pipelines capable of supporting large-scale training and post-training cycles while maintaining quality, consistency, and auditability.

We believe Innodata is well positioned to meet these requirements. Model Evaluation (“Evals”), Alignment, and Safety We believe that evaluation of model capabilities and safety (“evals”) are emerging as foundational layers of the AI technology stack, analogous to testing, security, and reliability engineering in traditional software systems. Unlike deterministic software, generative AI systems are probabilistic and context dependent.

Their behavior may vary across prompts, tasks, and deployment environments, and may change over time as models are updated or integrated with tools and new data sources. As a result, organizations increasingly require continuous evals to understand, measure, and manage model behavior throughout development and deployment.

These evals typically include: (i) capabilities evals that assess reasoning, knowledge, and task competence; (ii) alignment and safety evals that measure harmful behavior, misuse risk, and adherence to constraints; and (iii) regression evals designed to detect drift or degradation across model versions.

We believe this represents a durable and expanding market opportunity distinct from, but complementary to, data preparation and model training. From Output Scoring to Behavioral and Agentic Evals Early AI evaluation focused primarily on output correctness.

In contrast, today’s frontier systems - particularly agentic and tool-using systems - require behavioral and agentic evals that assess how models plan, reason, and act over time. These evals may examine reasoning coherence, tool selection and invocation, multi-step task execution, adherence to system instructions, and robustness under adversarial or ambiguous inputs.

This shift toward agentic evaluation materially increases the importance of structured human judgment, domain expertise, and scalable evaluation operations. We believe that the ability to measure not only what a model outputs, but how it arrives at those outputs, is increasingly central to deployment readiness and long-term safety.

Human-in-the-Loop Evals and Evidence for Trust As AI systems are deployed into regulated or high-stakes environments, customers increasingly require evidence that systems have been evaluated, documented, and monitored. This has driven demand for human-in-the-loop eval frameworks that combine expert judgment with automation to produce results that are interpretable, repeatable, and auditable.

Innodata’s evaluation programs emphasize rubricized scoring for consistency, subject-matter experts for high-risk domains, hybrid human-plus-automated evaluation pipelines, and longitudinal measurement to track regressions and improvements over time. We believe these capabilities position us to support emerging governance and regulatory expectations related to transparency, accountability, and risk management in AI systems.

Red Teaming, Adversarial Evals, and Safety Research AI safety has expanded to include misuse, exploitability, and unintended system behaviors - particularly as models are connected to retrieval systems, code execution environments, autonomous agents, and enterprise tools. Innodata conducts structured red teaming and adversarial evaluations to surface failure modes that are not observable through standard benchmarks.

These efforts include probing prompt-injection and jailbreak vulnerabilities, testing misuse scenarios involving retrieval-augmented generation, agent workflows, and tool use, identifying degradation under distribution shift, and supporting mitigation through targeted post-training datasets.

In parallel, we have expanded our cybersecurity capabilities as applied to LLMs and AI agents, including threat modeling for agent-based systems, assessment of data exfiltration and privilege-escalation risks, evaluation of secure tool invocation and sandboxing controls, and testing of monitoring and guardrail mechanisms designed to reduce exposure to adversarial attacks and enterprise security breaches. 5 Table of Contents We believe red teaming and adversarial evals are increasingly viewed as prerequisites for deployment rather than optional safeguards.

High-Risk Domains and Societal Safety As frontier AI capabilities advance, developers and governments have raised concerns about misuse in high-impact domains, including non-proliferation, chemical and biological risk, and large-scale misinformation.

Innodata supports mitigation efforts through domain-specific safety evals, targeted mitigation datasets, collaboration with academic and government-adjacent experts, and evaluation frameworks designed to preserve performance on legitimate use cases while reducing the risk of harmful behaviors or misuse. AI Model Deployment and Integration We believe that over the next decade, AI will be embedded across nearly all industries.

Innodata supports customers in operationalizing AI systems, including model customization, workflow integration, context engineering, and continuous quality assurance. Our platforms and services are designed to accommodate rapid innovation in model architectures and techniques, enabling customers to adopt new approaches without re-architecting their AI operations.