How Mediacorp brought observability to AI at national-media scale

Scaling AI content processing across millions of assets

Mediacorp is Singapore's national media network, engaging 99 percent of the population every week across its platforms. It is also the country's largest content creator, managing a catalogue of roughly 5 to 6 million images, 150,000 hours of video and 16,000 hours of audio, with volumes growing every month. For an organization of that scale, metadata is not a technical detail. It is what determines whether content reaches the right audience at the right moment.

Editors searching for archival footage, recommendation systems surfacing relevant programming and advertising systems targeting audiences all depend on accurate, machine-readable descriptions attached to every asset. Manual tagging was no longer feasible, and without automation the gap between the content Mediacorp produced and the content its teams and audiences could actually find kept widening.

As AI capability spread across the content pipelines, the models grew more sophisticated, progressing from trained baseline models to large language models and multimodal systems. That raised harder questions than whether the technology worked. Were the outputs accurate? How much compute and how many tokens were being consumed? Which models performed best for which content types? Mediacorp needed an integrated view of all three to balance performance against cost.

Turning Amazon CloudWatch into an AI observability layer

Mediacorp selected SoftwareOne through a competitive procurement process, choosing us for both technical depth and adaptability. Working alongside our partner AWS, the team applied the AWS Business Value Realization framework to define and track the outcomes that mattered to Mediacorp from the outset. Because the AI landscape shifted during the engagement, with new large language and multimodal models becoming available, SoftwareOne continually adjusted the solution architecture to reflect what

produced the best results. The observability layer was built on Amazon CloudWatch. Rather than using CloudWatch for infrastructure monitoring alone, SoftwareOne extended it to cover AI-specific signals including model accuracy, token consumption, latency and cost per workflow, making this one of the first known uses of Amazon CloudWatch applied to AI workloads.

The approach is deliberately non-prescriptive. Instead of applying a fixed set of predefined policies, SoftwareOne's machine learning engineers worked alongside Mediacorp's teams to instrument the existing pipelines, surface where accuracy and spend were drifting, and act on what the data showed. Confidence built quickly. Partway through the engagement, Mediacorp expanded the monitoring scope from four workflows to nine, a shift from cautious adoption to active investment.

The solution was structured around four pillars:

1. Infrastructure rightsizing, where Amazon CloudWatch metrics tracked CPU and memory usage across AWS Lambda and Amazon Elastic Container Service (across more than 8,900 Lambda invocations over 30 days, memory use on key functions averaged below 25 percent, with one function using just 8.5 percent of its allocation).
2. Model selection, where Amazon CloudWatch tracked performance, latency and token cost across models available through Amazon Bedrock (Anthropic Claude Sonnet), with dynamic model selection routing each task to the best-performing option;
3. Continuous accuracy monitoring and A/B testing, with F1 scores tracked per workflow;
4. Issue response, where AWS X-Ray gave engineers a structured path from alert to root cause through the service topology map, segment timeline and exception detail.

Dynamic model selection proved one of the most valuable capabilities. By tracking historical performance, cost and quality metrics across models simultaneously, the team could identify which model produced the best results for a specific task.

Sixty percent of entity-linking workloads were routed from a higher-cost model to a lower-cost alternative that performed equally well on those tasks, cutting token costs by 55 percent with no loss in output quality.

The same tracking data drove accuracy improvements. The named entity recognition workflow's F1 score, a measure of a model's precision and recall, registered at 0.80. That visibility prompted the team to build an improved model, which achieved 0.886.