AI for analyst workflows
SUNWEST BANK · AVP, AI/ML ENGINEER · OCTOBER 2025–PRESENT
My role is technical lead for enterprise AI across lending, underwriting, M&A, and compliance, with ownership from architecture through production.
The problem
Analysts need to turn large document collections and regulatory material into decisions they can defend. Retrieval and generation must support review, with evaluation at the level of the fields and decisions the workflow depends on.
What I built
For BSA/AML alert review, I led triage, structured extraction, retrieval-grounded classification against a regulatory rule library, and narrative generation with human review. Reviewer decisions are logged as labeled data for subsequent improvement.
- Alert triage & extraction
- Retrieve regulatory context
- Classify & draft narrative
- Human review & feedback
This is a conceptual workflow derived from my résumé, rather than a disclosure of internal infrastructure.
I also built and own Enclave, an M&A due diligence platform over deal rooms containing thousands of documents: OCR and layout parsing, embeddings, hybrid retrieval in Azure AI Search, multi-step RAG, and model routing by complexity, latency, and cost.
Evaluation and outcomes
| Workflow | Reported result |
|---|---|
| BSA/AML alert review | 4,000 analyst hours saved per year; 70% less review time |
| M&A research | 80% less research time per deal |
For underwriting model work, I benchmarked QLoRA-tuned Phi-3 against zero-shot, few-shot, and grounded GPT-4 baselines using field-level precision, recall, and schema validity. Those model comparisons are separate from the workflow time savings above.
Internal datasets, sample sizes, and the time-savings calculation methods are not public. These résumé-reported results are not independently verified by this website.
Engineering judgment
Human review is a boundary in the alert workflow. Model selection depends on complexity, latency, and cost; model quality is evaluated against baselines rather than assumed from model size. I co-authored a four-tier AI risk framework aligned to SR 11-7 / OCC 2011-12, covering evaluation, monitoring, sensitive-data handling, and human review.