Architected a plugin-based tool registry for UKG's internal FastMCP server. 15+ domain teams have registered 60+ tools that the AI assistant discovers at runtime, with no changes to the agent code.

• Designed a real-time customer feedback pipeline on Pub/Sub and BigQuery. It ingests 50K events a day, gives product teams same-day sentiment, and saves analysts about 6 hours a week of manual data pulls.
• Auto-instrumented 45+ microservices with OpenTelemetry and wrote a custom Go OTEL Collector that sends traces to Grafana and Arize and metrics to BigQuery and Splunk. Incident resolution got 40% faster and query times 53% faster.
• Built a custom OTEL processor for data privacy that batches JSON attribute values into Google DLP tables and rebuilds the spans after masking. Less PII exposure, and 60% off the DLP API bill.

Wrote the C and C++ firmware for a 4-in/4-out audio mixer, splitting a dual-core microcontroller so signal processing had one core and the console workload the other, and neither stalled. All 8 channels ran dropout-free.

• Sped up the JUCE DSP chains with decimation (lower sample rates) and FFT-based frequency analysis, taking a full analysis pass from 430 ms to 322 ms, 25% faster.
• Brought audio latency from 12 ms down to under 5 ms by tuning DMA and the RTOS kernel parameters.
• Built a Python and NumPy test harness that compares 200+ FFT captures to reference outputs on every build. It caught around 12 DSP regressions, and every fix shipped with a regression test.

Built an SMS tool in Java and Spring that let 200+ users chat with AI models from any basic phone, with Twilio handling inbound texts and each conversation kept in context across sessions.

• Developed a provider-agnostic LLM layer with LiteLLM across Claude, ChatGPT, Gemini, and Ollama, with failover and hot-swapping mid-conversation. Model switching latency dropped 56%.
• Replaced hand-built images and hand-configured GCP resources with Docker, Terraform, and a CI/CD pipeline that deploys on merge. Release setup time dropped 70%.