Three musts for running AI at network scale: A live discussion

"There is no way that a human engineer could actually execute at the level of scale and speed that we are seeing with these applications now."

Deploying AI-driven applications reliably, at speed, across a network that never stops changing is where most operators will encounter their hardest engineering problems.

On this week’s episode of Zero-Touch Live, Rakuten Symphony CMO Geoff Hollingworth spoke with Petrit Nahi, Chief Consultant for AI and Data at Rakuten Mobile, who has spent seven years building the data and automation systems that now run autonomously across Rakuten Mobile's live network. The journey traces back to PhD work he began more than two decades ago, when he focused on distributed multi-agent systems for dynamic network coverage that are not substantially different from what’s being implemented today.

📺 Watch the replay now below.

Non-negotiables for autonomous operations

Rakuten Mobile made a foundational decision at launch: centralize all network data and give ownership of that instrumentation to a single team. That decision, more than any model or algorithm, is what made autonomous operations possible, says Petrit.

Geoff and Petrit discussed the engineering realities that only become visible once AI applications start running:

• Data centralization is a prerequisite. Rakuten Mobile consolidated all probe and trace data under one team from day one in contrast to traditional operator organizations where these tools may be owned by separate departments, creating data separation that introduces challenges.
• Moving from network element to subscriber-centric data is a step change in platform requirements. Network element data is aggregated across thousands of elements while the evolution to subscriber-centric customer experience and performance data means tracking individual experience across an entire network, representing a many-fold increase in volume that compounds at scale.
• AI applications stress OSS systems before they take a single action. Because autonomous applications cannot assume network state is unchanged, they must fetch current state before every decision cycle. At Rakuten Mobile, that read load alone exposed the limits of existing downstream systems and required re-dimensioning.
• Re-dimensioning downstream systems is inherently complex. Cloud-native infrastructure made scaling easier in some layers though some network elements still required software upgrades that made adapting to AI-era request rates take significantly longer.
• Parallel autonomous applications must be coordinated. Energy saving, coverage compensation and fault response can conflict when running simultaneously, underscoring the importance of governance structures to manage those conflicts.

Petrit's closing point for any operator eager to scale AI deployments was to focus on three critical requirements: consolidate the data first, eliminate organizational silos around it and then ensure the platforms underneath are actually scalable enough to support what autonomous operations will demand from them.

‍