Solutions Architect Intern

Solutions Architects and their customers routinely sit on raw datasets, usage logs, cost exports, telemetry dumps, 10GB or more, with a real question behind them and no fast way to reach an answer. Getting from a raw file to a first chart and a written insight meant loading the data, writing analysis code by hand, and iterating on visualizations: a four-hour round trip on a good day, with no self-serve path for a non-technical stakeholder to ask a question of their own data.

I designed Lumina as a full-stack platform: React and TypeScript on the front end, a Java and Spring Boot API layer, and a backend built around Amazon Bedrock agents. Cognito handled authentication and role-based access with JWT tokens, so datasets stayed scoped to the tenant that uploaded them. S3 held everything across dedicated buckets for raw data, generated scripts, reports, and visualizations, and API Gateway routed requests into Spring Boot and into two purpose-built Lambda functions behind it. EC2 auto scaling behind an Application Load Balancer, fronted by CloudFront, kept the platform responsive under uneven traffic, holding 99.9% availability at sub-two-second response times.

One Lambda function handled visualization rendering: it packaged matplotlib, pandas, and seaborn into a Lambda layer, took a Bedrock agent's interpretation of a natural-language prompt, and executed the generated Python against the uploaded dataset (CSV, Excel, JSON, or Parquet), writing the resulting chart back to S3 behind a presigned URL. A second Lambda function ran a Bedrock agent with code-interpreter access to S3, reading a customer's file directly and generating equivalent Python and R analysis scripts on demand, complete with commentary, instead of a canned template. On the front end, a Monaco-based editor let a user read and adjust the generated code before rerunning it, with a Recharts-backed visualization layer rendering the output.

End to end, this took a four-hour manual analysis down to under two minutes on datasets of 10GB or more, and cut manual coding time on custom analysis requests by roughly 80%.

Before writing code I ran discovery sessions with the Solutions Architects who'd actually use the tool, to pin down what a first version needed to cover, and later led an architecture walkthrough for the wider team covering the Bedrock agent design, data flow, and security model as Lumina moved toward broader use inside the org.

Bedrock's generated code couldn't be trusted to run unmodified: some scripts referenced local file paths, called plt.show() (a no-op in a headless Lambda), or assumed a naming convention that didn't match the platform's own. I built a code-adaptation layer that rewrote file I/O to point at the already-loaded dataset and stripped unsupported plotting calls before the code touched matplotlib's runtime, mixing Bedrock-assisted rewriting with deterministic pattern matching for the parts that didn't need a model in the loop. Keeping the pipeline flexible enough to answer a real question, without being so open-ended that a bad prompt could crash the container, was the core design tension of the whole build.

Lumina made it clear that agent-based systems need as much engineering around the model as they do model calls: validating output, sandboxing execution, and giving the agent a narrow enough job that its failures stay contained. I came out of the summer with real experience turning a Bedrock agent from a proof of concept into something reliable enough to sit behind a production-shaped platform.