Software Engineering Intern · May 26-August 6, 2026
At New York Life, I focused on turning agent concepts into demonstrable, testable systems—from multi-channel claims assistance to knowledge and developer workflows.
Primary focus
Applied AI engineering with careful trust, access, and evaluation boundaries.
Core work
The language here is deliberate: these systems were built and demonstrated or pilot-tested. They are not presented as production deployments or verified business impact.
01
I independently designed an asynchronous claims-assistance prototype for web, phone, and SMS. The web and phone demo covered public questions, service procedures, authentication transitions, structured claim intake, form submission, and document upload routing and prefill.
Because production claims and client data were unavailable, I researched service contracts and built mock integrations plus an administrative testing interface.
02
The assistant used workflow-encoded skills, public-source citations in the web experience, user-scoped information, domain boundaries, and modality-aware behavior. I evaluated it with four claims-stage personas and LLM-as-judge testing.
Tool failures were designed to degrade safely, with human escalation in web and phone and a graceful way to complete calls.
03
As a side project, I built connectors over enterprise sources I was authorized to access and turned their documents into linked topic pages. A daily freshness flow identified new material, summarized it, linked back to primary sources, and refreshed relevant topics.
The goal was richer, traceable answers across enterprise context—not a replacement for source systems or their permissions.
04
I built a local AI-searchable knowledge base and read-only discovery tools for a large low-code rules corpus. I packaged the workflow and trained three developers to incorporate it into their own work.
Public details are intentionally limited to the workflow and training; internal system mechanics and corpus figures are omitted.
Capstone pilot
I proposed the skills approach and handled key developer-enablement pieces: integrations, local preview, and codifying the workflow. We tested it across several coding-agent environments.
Step 1
A design agent combined a design, work ticket, and existing-code context into an implementation plan.
Step 2
A coding agent implemented, tested, and opened the work in a local preview.
Step 3
A verification loop compared the preview with the design and iterated on visual differences.
Controlled test observations
~30 min
Simple compositions, with 95-100% measured visual accuracy.
60-70 min
More difficult compositions in the pilot tests.
These are bounded test results, not a claim of released productivity gains, organization-wide adoption, ROI, or cost savings.
Additional contributions
Collaborated on a Microsoft Copilot agent for requirements summaries and scoped user stories using approved enterprise context. Deeper source-repository-aware planning remained incomplete.
Helped architect a pipeline for difficult form and document cases. I am not claiming production deployment or an outcome metric.
Created a practical AI and skills kickstart guide and trained colleagues across business and engineering roles.
Evidence boundary
Every capability described here was built, demonstrated, directly completed, or tested in a pilot. Exploratory avatar and representative-copilot work is intentionally omitted from the delivered work. This page does not disclose private data or system details, and it does not claim production deployment, customer outcomes, staffing impact, cost savings, or ROI.