Production work describes my contribution and its scope. The personal project shows how I apply the same engineering principles on my own.
Federal healthcare · Production work
~2 hours → <1 minute
From a two-hour wait to under a minute
Tech Lead · Reporting service and upstream publishing integration
Problem
A set of reports on a national healthcare platform ran on a roughly two-hour cron cycle. Users waited for the next cycle after submitting a file.
My approach
I replaced polling with SQS messages that triggered report generation when a file was ready. When the upstream team had no capacity, I implemented the publishing integration in their codebase, with their engineers reviewing and approving it. I checked correctness against the agency's report requirements.
Outcome & reflection
Reports came back in under a minute. The bottleneck was the trigger, not compute: changing when the work started removed the wait.
TypeScriptNode.jsAWS SQSPostgreSQL
Public health · Production work
Manual checks → daily evaluation
Making AI quality measurable, every day
Tech Lead · Built the evaluation pipeline, not the assistant
Problem
Manual spot checks gave the team no consistent way to assess whether changes to a public health AI assistant improved its answers.
My approach
I built a daily pipeline using a fixed set of questions and reference answers, LLM-as-judge scoring, separate retrieval evaluation, and assessment of production traces. Node.js ran on AWS Lambda, with Langfuse for tracing and scoring and Weaviate as the vector store. Scheduled runs captured drift even when nobody deployed.
Outcome & reflection
The team gained a repeatable daily quality signal. A next step would be calibrating the judge against subject matter experts to measure its own error rate. This was scheduled evaluation, not a real-time guardrail.
Personal project · Application design and implementation
Problem
Keeping feeding logs in a chat assistant made it difficult to retrieve reliable totals. A conversation alone was not a dependable system of record.
My approach
I built an app that accepts natural-language questions and returns averages and charts. The model interprets the request; stored data and database calculations determine the result.
Outcome & reflection
A practical application of a principle I use in AI systems: keep arithmetic and durable records outside the model. This is a personal project, separate from my production work.