FCE INTERVIEW COMMAND CENTER

"I'm Vitus Dzekedze — a multi-cloud AI architect and data engineer with over 15 years building production systems in highly regulated environments, including JPMorgan Chase, Bank of Hawaii, and the Alamo Colleges District here in San Antonio.

What draws me specifically to FCE is the combination of two things I know well: health and welfare benefit administration and Google Cloud AI infrastructure. You're migrating to GCP and building AI from the ground up — that's exactly the kind of greenfield challenge I thrive in.

I'm also one of the few engineers in this market who is Anthropic-certified — specifically in Claude Code, AI Fluency, and the Claude implementation framework. The JD calls out LLM deployment, RAG pipelines, and responsible AI in regulated environments. Those aren't aspirations for me — they're my recent work."

"My AI/ML work spans two tracks: enterprise data pipelines and generative AI deployment. At JPMorgan, I architected pipelines processing 2+ million daily financial transactions with 99.9% uptime — that's the scale and reliability mindset I bring to ML infrastructure.

On the generative AI side, I've built LLM-powered Q&A systems, document summarization agents, and RAG pipelines using both Claude and Gemini APIs. I use Vertex AI on GCP for deployment and BigQuery for the data layer — which lines up directly with your platform stack.

I'm also Anthropic-certified, which means I understand not just how to deploy LLMs, but how to do it with explainability, auditability, and HIPAA alignment built in from day one — not bolted on later."

"For a claims summarization system, I'd architect a RAG pipeline on GCP: raw claims documents are ingested via Cloud Storage, embedded with Vertex AI text embeddings, indexed in a vector store like Vertex AI Matching Engine or AlloyDB with pgvector, then surfaced through a Claude or Gemini-backed LLM with context-aware summarization.

For member eligibility Q&A, the pattern is similar but real-time — a member asks 'am I covered for this procedure?' The system retrieves their plan document segments via semantic search and generates a plain-language response with cited sources, so every answer is auditable.

The HIPAA layer runs throughout: PII redaction at ingestion, role-based access controls via GCP IAM, encryption at rest and in transit, and full audit logging of every query-response pair. That audit trail is also your SOC 2 evidence artifact."

"I work across all three major clouds but have deep hands-on experience with GCP's AI stack — specifically Vertex AI Pipelines for end-to-end ML workflows, BigQuery ML for in-warehouse model training, Dataflow for streaming data processing, and Cloud Composer for orchestration.

For LLM work, I've integrated both the Gemini API and Claude via GCP endpoints. I use Cloud Functions for lightweight serverless inference triggers and Cloud Run for containerized API deployments.

I should be transparent: I hold Azure certifications (AZ-900, DP-900) and have Google Analytics certified, but I don't yet hold the GCP Professional ML Engineer cert. However, my production experience with the actual services — Vertex AI, BigQuery, GCS — is deep and current. I'm happy to demonstrate that technically."

"Security-by-design, not security-by-retrofit. I approach HIPAA compliance at five layers: data classification at ingestion (PHI tagged and isolated), PII redaction before any LLM call, role-based access controls via IAM with least-privilege principles, encryption at rest and in transit using GCP's built-in KMS, and complete audit logging of every data access event.

For AI specifically, I implement secure prompt handling — ensuring member PII isn't included in prompt context unnecessarily — and model access controls so only authorized service accounts can invoke LLM endpoints. The audit trail I build also doubles as your SOC 1 and SOC 2 evidence package.

I've done this at JPMorgan under SOX, and I've architected HIPAA-compliant healthcare data systems in my consulting work. Regulated environments aren't a constraint I work around — they're a design pattern I build from."

"I build MLOps with the same rigor as software engineering. My standard stack: GitHub Actions for CI/CD triggers, Vertex AI Pipelines or Kubeflow for orchestrated training and deployment, and model versioning via Vertex AI Model Registry. For monitoring, I set up drift detection on input distributions and output confidence scores, with automated alerts when models degrade.

At Vi-2s-Dk, I built automated MLOps pipelines that reduced incident response time by 70% — that was through proactive monitoring with Cloud Monitoring dashboards and automated rollback triggers when SLOs were violated.

For FCE's context, I'd also build in compliance checkpoints — automated model documentation for each version, bias testing reports, and explainability outputs — so you have audit-ready artifacts for every model in production."

"At Alamo Colleges, I led digital transformation for 60,000+ students across five campuses. I had to regularly explain data architecture decisions to finance directors and academic deans who had zero technical background. I learned to lead with outcomes — not 'we're implementing an ETL pipeline,' but 'you'll be able to see real-time enrollment and financial aid status in one dashboard, reducing processing time from 3 days to 4 hours.'

At UWM, I also produced over 500 technical tutorials designed to make complex concepts accessible — AI, Python, DevOps — for a range of audiences from junior developers to C-suite leaders. Communication clarity is something I've built as a deliberate skill, not assumed as a byproduct of technical knowledge.

For this role, working with the CFO's office, I'd apply the same principle: translate AI capabilities into cost reduction, compliance risk reduction, and throughput metrics that leadership can act on."

"At JPMorgan Chase, I architected and maintained data integration pipelines processing over 2 million daily financial transactions with 99.9% uptime. This wasn't a project — it was a live system where downtime had direct regulatory and financial consequences. I provided 24/7 production support and held strict SOX audit compliance throughout.

The lesson I carry from that environment: mission-critical AI systems require the same discipline as mission-critical financial systems. Monitoring, incident response, audit trails, and rollback procedures aren't optional extras — they're table stakes. That's the operational mindset I'd bring to FCE's AI build."

"TPAs handle high volumes of repetitive, document-heavy work — eligibility verification, claims processing, member communications, compliance reporting. These are exactly the workflows where well-designed AI creates measurable ROI.

Specifically for FCE, I see three high-impact opportunities: First, claims summarization — reducing the manual review time per claim by 60-80% using LLM-powered extraction and summarization. Second, member eligibility Q&A — an AI chatbot that handles the top 80% of member inquiries without routing to a human agent. Third, document intelligence — automating SCA/DBRA compliance checks across contract documents, flagging gaps before they become audit findings.

Each of these reduces administrative burden, improves member experience, and frees your operations team to focus on complex edge cases and relationship management. That's AI as a force multiplier — not a replacement, but an upgrade to your existing team's capacity."

"Governance before deployment, always. For every AI system I build in a regulated environment, I establish three things before anything goes to production: a model card documenting the system's purpose, training data, limitations, and risk profile; a data lineage map showing exactly what data flows through the system and where PHI is handled; and a human-in-the-loop escalation path for edge cases the model shouldn't handle autonomously.

For FCE specifically, this means any AI output that affects member benefits decisions has a human review checkpoint. The AI accelerates the workflow — it doesn't replace the accountable human. That's the responsible AI principle the Anthropic Claude framework builds around, and it's the standard I implement in practice.

The result is an AI system that strengthens your SOC 2 posture rather than threatening it — with complete audit trails and human accountability at every decision point."