Selectedwork.

// The problem

Give every employee an assistant over Slack, Jira, Confluence and a curated know-how base, with intelligent tools, that keeps data in-house, respects permissions, and only pays for GPUs while someone is using it.

// My approach

Scale-to-zero vLLM on GKE L4 and Qdrant hybrid search with ACL filters inside the query. A Pydantic AI agent streams over FastAPI SSE, deployed or destroyed with one OpenTofu, Helm and Ansible command.

// The problem

Hundreds of models, all running from notebooks — slow, brittle, no audit trail, no reproducibility.

// My approach

PySpark + Airflow platform driven by metadata config. Imbalance correction, parallel training, versioning, scoring. Every run reproducible from logged config. Internal teams add or modify clients from a UI — no engineering involvement.

// The problem

Alteryx / MySQL couldn't keep up. ETL ran for hours. No visibility. Engineering bottleneck for every new pipeline.

// My approach

Metadata-driven GCP architecture: BigQuery + Cloud Composer + Dataproc + PySpark, all IaC'd with Terraform. Analysts add pipelines via config — no engineering tickets. Incremental migration, production never stopped.

// The problem

Engineering owned all ops visibility. Business teams filed tickets to see their own data — slow, frustrating, trust-eroding.

// My approach

React + App Engine app pulling DAGs, BigQuery state, MySQL metadata, and logs into one view. Real-time updates, structured search, one-click validation, rerun / skip / override without engineering tickets.

// The problem

ROI reports: 30 min to hours, formatting errors. QBR decks: full day of slide assembly. Account managers buried in formatting.

// My approach

Django + React + BigQuery. ROI reports on-demand in seconds. QBR engine reads analytics tables and writes branded PowerPoint slides — typography, charts, narrative — programmatically.