Emitrr — Speech Pipeline & LLM Systems
Apr 2025 – Jan 2026
30% off transcription latency, a rebuilt TTS path, and an agent that learns what to do when a call falls apart.
AI / ML / GenAI Engineer
Most of what I build listens for a while before it decides anything.
At Emitrr I shaved a third off the latency of a live voice pipeline — the kind of number that only matters because someone was mid-sentence when it did. Outside of work, I build small experiments that poke at the edges of these systems: a narrator that never quite explains itself, a classifier that has to explain everything, a model asked to find a few pixels of bleeding in fifty thousand frames.
A rough shape of what I do
Production speech and LLM systems by day. Interpretable ML, agentic tooling, and the occasional game designed to make a language model slip, by whatever hours are left.
Non-Invasive Blood Glucose Estimation
Published research
A model that reads blood glucose out of a light-absorption waveform — one version accurate, one version an equation a person can actually read.
Signal lens
Switch the lens to see how the same portfolio changes depending on the problem it's solving.
Speech systems
Latency focus
350 ms
Accuracy signal
96%
Core question
Endpoint tuning and call-state recovery
System flow
Speech pipeline
Input
Live audio
Model
STT + endpoint logic
Decision
Turn boundary + latency correction
Output
Clean transcript / action