Selected systems

ML and AI projects

Selected systems Adrian Zevenster has designed, built, or operated across machine learning, AI, cloud infrastructure, and data engineering.

Featured project work

  • RedStrings Secure Generative Media Platform: Terraform, Cloud Run, IAP, Vertex AI, scheduled Drive intake, independent approval, and versioned asset provenance.
  • KYC Document Verification System: source-aware utility-bill, bank-statement, and financial-report workflows connected to Java 21, Gemini, Document AI, Kafka, Spanner, ClickHouse, and human review.
  • Utility Bill & Electricity Verification: provider-aware classification, weighted customer validation, expiry and forgery checks, external validation, confidence scoring, and regression coverage.
  • Gemini Entity Extraction Regression Framework: 19 golden cases and 17 suite configurations with fixture, replay, live-model, prompt-drift, and field-level release gates.
  • F1 Driver Intelligence: real-time telemetry and strategy guidance through specialist ML agents.
  • Teradata / ODI to BigQuery Migration Accelerator: live ODI discovery, plan-bound replay and refactor, signed private runners, deterministic and Vertex-assisted conversion, exact-commit Dataform promotion, and closed-loop reconciliation.
  • Floriculture Data & Forecasting Platform: validated Drive ingestion, source-specific Excel parsing, BigQuery and Dataform data products, Terraform-managed Google Cloud foundations, and nine-level hierarchical forecasts on Vertex AI.
  • LLM-Powered Data Analysis Platform: 33 schema-validated tools, deterministic and LLM planning, FAISS/BM25 retrieval, supervised ML, SHAP and conformal uncertainty, semantic memory, streaming FastAPI, and React.
  • CareLog Medical Assistant: multimodal health-data intake, human-confirmed evidence, tenant-scoped PostgreSQL/pgvector retrieval, citation-bound reports, FHIR R4 export, and privacy-preserving audit controls.