Marina
Gulakova-Hollingshead Production LLM tooling and event-driven backends.

The feature work is Kotlin and Angular over GCP, Snowflake and Postgres, supporting enterprise financial processes. The part I’d point at first is the tooling.

The evidence-based debugging skill I wrote constrains an agentic production investigation to sources it can actually cite: a log line, a metric, an APM span, a database row, file bytes, source at a named commit. One guided workflow correlates Kafka events, Snowflake and Postgres state, Dagster runs, GCP logs and Datadog metrics to get there.

Hypotheses have to be labelled as such, and paired with the query that would falsify them. From the evidence-based debugging skill

• Cross-service investigations went from days to a few hours, on a platform whose data flows were largely undocumented. The skill maps the flow as it goes, so each investigation leaves behind the documentation the system never had.
• Named the failure modes and encoded them. Correlation read as causation. A root cause generalized from a five-row sample. Runtime behavior inferred from reading code. A stale local checkout quoted as deployed code.
• Absence of evidence has to be reported as an explicit negative, naming the retention window it covers. “I found nothing” and “nothing exists” are different answers.
• Designed and proposed a triage service for proxy failures. It reads off Kafka dead-letter queues, uses Gemini Vertex AI and the Claude Code debugging skills to isolate root cause, and opens a Jira ticket with a proposed fix attached — scoped to run against any of the twenty or so backend services in the organization, not only the seven my team owns.
• Wrote the steering and governance docs for several backend services, so agentic tools generate production-ready code instead of plausible-looking code. Merge request review skills run the first audit pass against team standards before a human opens the diff.
• Wired Jira, Datadog, Snowflake and Dagster MCP servers into the team’s development environment.

One of several technical consultants on the transition. My piece was the data pipeline for the meter fleet: roughly 173,000 devices across utility customers in emerging markets, each emitting a heartbeat every fifteen minutes, which comes to about 16.5 million readings a day.

• It reads meter data straight from the cloud and writes it as Parquet in S3, then generates the daily and monthly consumption and billing reports the utilities run on. The path it replaced staged everything through a database first.
• Built the customer-facing APIs for retrieving those reports.
• Added observability and current debugging patterns to legacy IoT services, which brought time to resolution down on code nobody on the new team had written.

• Python backend services behind deployment and account management across the same 173,000-meter fleet.
• Built the reporting pipeline that read ingested meter data out of the database and generated the consumption and billing reports delivered to utility customers.
• Elixir and Phoenix features for real-time meter data processing and customer usage analytics.
• Vue.js components for the meter monitoring and customer management dashboards.
• Maintained and upgraded the APIs for device provisioning, configuration management and remote diagnostics, the parts field teams depend on.
• Added custom CI/CD pipelines and rewrote the slow queries that were setting response times.

Co-designed and built, from nothing, the backend connecting Nike’s internal applications to the third-party platforms its marketing organization ran on — Sprinklr, Airtable and others. Java and Spring Boot microservices, a React frontend, both SQL and NoSQL stores, around 1.5 million transactions a day.

• Led a team of eight. Set the technical standards, made the architecture calls, owned platform features from prioritization through release, and mentored the junior engineers.
• Established Docker deployments to AWS through Jenkins, which gave the team real continuous delivery.
• Designed REST APIs to Nike’s architectural standards and integrated the platform with a long list of internal systems.
• Ran root-cause analysis on production incidents and built the error reporting that made recurring failures visible instead of anecdotal.
• Modernized legacy AngularJS applications onto a Java Spring backend using Alfresco, cutting server response times.
• Nike MVP award, FY20 Q1.