Backend.
Platform.
Applied AI.

A customer needed Cisco CallManager records displayed as complete calls. Records arrived across one-minute files, with links from media to segments to call IDs. No message established that every component had arrived.

I built three asynchronous queues: wait for files to settle, assemble call components, then batch and reliably deliver the results. A customer-selected timeout started with the call-start record; late records were treated as a new call. The completion policy made the uncertainty explicit.

A vendor PDF renderer leaked memory until the web server crashed. Live memory debugging showed unmanaged growth, while managed memory was not the source. I traced the problem to the vendor’s .NET wrapper allocating memory improperly for its C++ engine.

I rewrote the wrapper while retaining the rendering engine. The replacement completely cured the memory leak.

Serializing the full character and inventory state to JSON caused visible gameplay stutters. I captured a binary snapshot in memory, then serialized that snapshot on a background thread before cloud persistence.

This introduced the game client’s first asynchronous path. The stutters disappeared completely.

I built an LLM-assisted tool to translate complex regex log parsers into Lua LPeg. With no real samples available, the model also generated test messages. I first ran each candidate message through the original regex, then used accepted messages to compare the translated parser’s results and capture values.

Executable checks challenged both the generated tests and the generated code. Roughly 80 rules were translated. Passing examples was evidence of correctness, not proof of equivalence for every input.