AI Systems & Evaluation

Eric Voss

I build document processing and agentic AI systems for regulated industries, and the evaluation practices that show whether they work.

I'm a Minnesota-based data scientist at Securian Financial, where I design intelligent document processing pipelines, architect agentic AI systems on Amazon Bedrock, and set the standards for how AI gets evaluated before anyone depends on it. All of it in an industry where PII and PHI handling, HIPAA, and model governance are constraints on the design rather than paperwork added at the end. My career spans more than a decade across application development, analytics, and AI/ML in the financial services industry.

Although my professional focus these days is mostly on practical topics like document AI, agent architecture, and evaluation, I also like to dive into more off-the-wall research areas while exercising my "traditional" data science and statistical modeling muscles. Whatever comes of that ends up in posts — full write-ups when the work had a result worth showing, shorter notes when it didn't, and a few things still being figured out in the open.

LinkedIn GitHub Email

What I work on

More about me

Intelligent document processing Designing and building AI pipelines that turn applications, policy documents, and correspondence into structured data — classification, extraction, validation, and the confidence thresholds that decide what a person still has to review before anything downstream trusts it.

Agentic AI architecture Architecting and building agentic systems on AWS, primarily Amazon Bedrock AgentCore — tool and memory design, orchestration across steps that can fail, and the identity, isolation, and guardrail boundaries an agent needs before it is allowed near PII or PHI.

Evaluation & enterprise AI standards Systematic evaluation of AI systems: test sets, scoring, and regression checks that produce evidence rather than a demo — and the standards and practices that define how AI gets built, reviewed, and operated across the company.