I build things
that listen.
Software Engineering Intern at Hemut / Summer 2026
Explore the workShipment reference muted / Original timing
Agent: Thanks for calling Sunview Logistics. How can I help you today? Caller: Hi, can I check up on a load status? Agent: Yeah, I can do that. What's your load or reference number? Caller: [Load reference muted.] Agent: Alright, one moment. Right, [load reference muted] is delivered.
Automatic transcript, edited to mark redactions. 19-second excerpt. Shipment reference muted; original pace and pauses preserved. The source is a 48 kHz Opus recording.
I tested speech models on the situations our voice agents actually encounter. A fast transcription isn't much help if it misses the thing a caller needs.
Inside the benchmark40 synthesized domain clips; a 7-clip name subset for keyterms
Three reported baselines; one keyterm projection. Local and hosted runtimes differ.
Historical reported summary from open-weight-stt-findings.md, not a fresh run. WER is a fraction; latency is the reported median in milliseconds. Whisper and Parakeet ran locally on an M4 GPU; Deepgram used a hosted API, so timing includes different compute and network conditions, not live-turn latency. The keyterm row is the same Deepgram system: full-corpus WER is projected from seven name clips, and 204 ms is inherited from the baseline, not remeasured. The separate 240-row probe pools six audio conditions and is not this 8 kHz summary.
Production systems.
Independent experiments.
Voice agents for freight calls, with work spanning the conversation, the backend, and how both reach production.
A logistics copilot that finds the tools a question needs and coordinates answers across them.
A speech-recognition benchmark built around freight calls, including a scoring correction that changed my model comparison.
A Chrome extension that checks article claims against sources and places the results beside the text.
I chose Perplexity because I preferred its search at the time. Bring your own key kept me from taking on ongoing API bills, and the extension did not need a backend of my own.
A chat app that reuses similar answers, routes simpler questions to a smaller model, and shows modeled energy savings.
I handled the engineering and worked with one designer. The app combined a chat interface, semantic caching, model selection, and an energy dashboard.
I'm DongYeop Lee, or DY. I spent Summer 2026 at Hemut working on voice agents and the backend behind them. That meant moving between conversations, code, and the tests that were supposed to tell me whether any of it worked.
Seeing issues in Converse sent me back to my own benchmark. I had already reached a conclusion, but the scorer needed another look. I care about that part of building just as much as getting the first version running.
This site is a project, too. I'm building mine first so my designer partner and I can learn from it before we build hers. Veracity and EcoPrompt are here for the same reason: I wanted to make the idea real and see what happened.
“I called my own agent. I got tired of waiting.”Why I went after the latency.
Voice agents
Backend infrastructure
AI evaluations
NLP at IDX Exchange
A browser extension
A hackathon with a designer
