Case Study: Scaling 5G Platform Under Conflicting Priorities

At Verizon, I owned product strategy and execution for core platform capabilities powering 5G-enabled digital customer experiences. The platform supported multiple downstream teams and millions of end users, making reliability, performance, and speed critical to both user experience and business outcomes.

However, as adoption of 5G services accelerated, the platform began to experience increasing friction:

• Rising support tickets tied to inconsistent system behavior 
• Growing backlog of competing feature requests from multiple teams 
• Slower release cycles due to cross-team dependencies and unclear prioritization 

The core challenge was not just technical, it was product ambiguity at scale.

Multiple teams were requesting high-priority features simultaneously, often with conflicting timelines and unclear business impact. Prioritization was reactive, driven by urgency rather than structured evaluation.

This led to:

• Fragmented roadmap decisions 
• Increasing operational overhead 
• Degradation in customer experience 

The key question became:

How do you prioritize effectively in a high-scale platform where every request is “urgent”?

Instead of evaluating requests individually, I shifted the framework to focus on:

• Customer impact (user friction, failure rates) 
• Business impact (revenue potential, adoption) 
• System impact (reliability, scalability) 

This created a shared language across stakeholders and reduced subjective prioritization.

Rather than relying on stakeholder input alone, I partnered with data teams to analyze:

• User interaction patterns 
• Failure points across workflows 
• Support ticket trends 

This surfaced a key insight:

A small number of systemic friction points were driving a disproportionate share of user issues.

Instead of continuing feature-by-feature prioritization, I:

• Re-prioritized the roadmap around high-impact platform improvements 
• Deferred lower-impact feature requests 
• Aligned stakeholders on long-term platform health vs short-term wins 

This required navigating tradeoffs and pushing back on competing priorities.

I worked closely with:

• Engineering to define scalable solutions 
• Operations to manage dependencies 
• Stakeholders to maintain alignment and transparency 

We also introduced more consistent Agile practices to improve delivery predictability.

• Prioritized system reliability over incremental feature delivery 
• Shifted from reactive prioritization → data-driven roadmap decisions 
• Invested in platform improvements that reduced downstream complexity 

These were not universally popular decisions, but they were necessary to stabilize the system.

• Reduced support tickets by 54% by addressing core friction points 
• Improved time-to-market by 30% through better prioritization and delivery practices 
• Enabled new revenue-generating capabilities with $20M projected lifetime value 
• Supported platform features used by millions of customers 

1. Not all urgency is equal

               In platform environments, stakeholder urgency often masks underlying system issues. True impact c        comes from identifying and solving root causes.

2. Data creates alignment where opinions cannot

              Behavioral data was critical in shifting conversations from subjective prioritization to objective decision-making.

3. Great PM work is often about saying no

              Driving meaningful impact required deprioritizing lower-value work and aligning teams around fewer, higher-impact initiatives.

This experience shaped how I think about:

Platform product strategy

Tradeoffs between short-term delivery and long-term scalability

Using data to drive alignment in complex systems

It also directly informs my current focus on AI-driven systems, where similar challenges exist in evaluating performance, prioritizing improvements, and balancing experimentation with reliability.

As AI-driven workflows and agent-based systems became more accessible, I began experimenting with building lightweight agents to automate multi-step tasks such as information retrieval, synthesis, and structured output generation.

While initial prototypes demonstrated strong potential, I quickly observed a gap between:

• Impressive demo performance 
• Consistent real-world reliability 

This highlighted a core product challenge:

How do you evaluate and improve AI agents beyond surface-level outputs to ensure they are reliable in real-world use?

Early iterations of agent workflows showed:

• Inconsistent output quality across similar inputs 
• High variance in response time (latency) 
• Failure modes that were difficult to predict or diagnose 
• Lack of clear metrics to evaluate performance systematically 

Traditional product metrics were insufficient for capturing:

• Output correctness 
• Consistency across runs 
• Degradation under edge cases 

The key challenge became:

 Defining a structured evaluation framework for systems that are inherently probabilistic.

I established a set of core evaluation dimensions to assess agent performance:

• Output Quality: Accuracy, relevance, and completeness of responses 
• Consistency: Stability of outputs across repeated runs with similar inputs 
• Latency: Time taken to generate responses, especially for multi-step workflows 
• Failure Modes: Types and frequency of breakdowns (hallucinations, incomplete outputs, tool failures) 

This created a standardized way to compare iterations and identify tradeoffs.

To move beyond ad hoc testing, I created:

• Representative task sets covering common and edge-case scenarios 
• Repeatable inputs to test consistency 
• Scenarios with increasing complexity (single-step → multi-step workflows) 

This enabled controlled evaluation rather than anecdotal assessment.

I implemented lightweight logging to capture:

• Intermediate reasoning steps 
• Tool usage patterns 
• Points of failure or degradation 

This provided visibility into why failures were occurring, not just what the outputs were.

Using insights from evaluation, I made targeted improvements:

• Adjusted prompt structures to reduce ambiguity 
• Introduced constraints and guardrails for more deterministic outputs 
• Optimized workflow steps to reduce latency and unnecessary calls 
• Refined fallback behaviors for failure scenarios 

Each iteration was measured against the evaluation framework to ensure improvements were real and not anecdotal.

• Prioritized consistency over peak output quality, recognizing that reliability drives user trust 
• Invested in evaluation infrastructure early, rather than scaling unstable workflows 
• Focused on failure mode analysis, treating breakdowns as product signals rather than edge cases 

• Improved output consistency across repeated runs, reducing variability in results 
• Reduced latency by optimizing workflow steps and eliminating redundant operations 
• Identified and mitigated key failure modes, leading to more predictable system behavior 
• Established a reusable evaluation framework for future AI-driven workflows 

1. AI systems require new product thinking

        Unlike deterministic systems, success is not just about correctness, but about managing variability and uncertainty.

2. Evaluation is a product problem, not just a technical one

         Defining the right metrics and frameworks is critical to making meaningful progress.

3. Reliability drives adoption

         Users are more sensitive to inconsistency than occasional errors. Predictability builds trust.

This work directly informs how I think about:

Building and scaling AI-powered products

Evaluating agent performance in production environments

Balancing experimentation with reliability and user trust

As AI becomes more integrated into core product experiences, I’m particularly interested in solving challenges around:

Agent orchestration

Performance evaluation at scale

Designing systems that are both adaptive and dependable