HamzaGabajiwala

Software Development Engineer I · Dublin, Ireland

• Built Accelerated Audience Activation end-to-end — Spark + Avro + bucketed user partitioning + backward-compatible named-parameter scoring app — that cut new-segment activation latency from 0–0 hours to ~4 hours.
• Built the GenAI keyword-expansion DAG (Airflow + EMR Serverless + Amazon Bedrock, currently Claude Sonnet 4.5) — scaled LLM concurrency from 0 to 0 threads to saturate the 0-RPM model quota and shipped inference-profile reuse that stopped Bedrock's 0-profile per-region cap from killing hourly runs.
• Shipped 3-level real-time re-engagement targeting in Flink (Line → Package → Campaign) — extended the segment cache with TLongObjectMap package/order indexes, added a DSP line mapping cache loaded from S3, and feature-flagged the rollout so it activates only when the new rule types are exposed.
• Designed a segment-reprocessing system covering Yahoo DSP's 0K-segment audience catalog — daily health-check DAG with Slack alerting, a forward-compatible write_target toggle for the upcoming OpenSearch → S3 cutover, and remediation playbooks that resolved live customer incidents (traced 0 segments, restored 0 for a major travel advertiser — ~0M users brought back).
• Migrated 4 production scoring systems — upgraded data access layers, moved from EMR v6 → v7, replaced Glue catalog reads with direct S3, and migrated monitoring to Chronosphere via OpenTelemetry — cutting batch-scoring cost 0% and unlocking sub-minute alerting.
• Led the team's adoption of agentic development tooling — consolidated 4 product repos under a shared submodule layout, published a Claude Code plugin marketplace with 7+ shared skills, and ran weekly knowledge-sharing for the Dublin team.

Software Developer · Mumbai, India

• Migrated 0+ test cases from Selenium to WebDriver in 0 days — manual regression to automated nightly runs.
• Cut data-collection downtime by 0% with REST API ingestion, enabling same-day reporting.