Skip to main content
Let’s talk
All experience

Agentic AI for telephony at Apple

Role
Software Engineer Intern, Apple
When
Mar 2026 to Sep 2026
Where
Cork
Result
Incident triage time cut by ~90%
  • Python
  • LangGraph
  • MCP
  • Kubernetes
  • Kafka
  • Redis
  • Docker
  • Splunk
Interns gathered around the rainbow Apple sculpture in the atrium of Apple's Cork campus.
With the intern cohort at Apple, Cork.

Context

Apple's Global Voice Services (Telephony) team keeps voice infrastructure running at global scale. When a voice incident occurs, engineers historically had to triage it by hand, digging through telephony systems, correlating signals/call traces, and working out the root cause under time pressure. I joined the team in Cork to automate that.

What I built

The core of my work was an AI Root Cause Analysis Engine that automates the analysis and triage of voice incidents:

  • A LangGraph state graph coordinating multiple AI agents with parallel execution, conditional routing, and retries, analysing up to 14 calls in parallel.
  • Guardrailed MCP servers deployed on Kubernetes pods, exposing telephony systems as agent-callable tools, monitored through an observability app for debugging and health.
  • A PII tokenisation pipeline that sanitises sensitive data before anything reaches an LLM, with mappings stored in Redis under a TTL policy.
  • A Kafka consumer that ingests ServiceNow incidents in real time, forwarding relevant cases to the RCA engine.
  • Production hardening: mTLS authentication, rate limiting, 500+ automated tests, Splunk observability, and a CI/CD pipeline.

I led the project as tech lead. Beyond the code, I presented the project charter to senior managers and stakeholders to align on scope and delivery milestones, and collaborated with engineers across multiple time zones.

Impact

The engine reduced incident triage time by roughly 90%, turning a slow, manual investigation into an automated first pass that engineers can act on immediately.

What I learned

Resilience, mostly. Getting the application approved for production meant working through a long chain of bureaucratic processes, and the skill that actually mattered was keeping a cool head and staying fixed on the objective rather than the obstacle in front of it.

Technically, it was an education in what production really demands: the depth of security, access and authorisation control involved, and how much of it ultimately comes down to identity.

The bigger lesson was about people. The amount of unblocking I got purely from reaching out (on Slack, in person, in casual conversation with other interns) was genuinely surprising. Communication turned out to be the highest-leverage tool I had.

Related