Rupesh I build real-time voice AI systems where model behavior, latency, and robustness matter as much as raw capability. As a Founding AI Engineer, I specialize in low-latency streaming inference architectures for conversational systems, optimizing throughput, analyzing latency–accuracy tradeoffs, and designing reliability primitives that prevent silent failure in production inference pipelines. My work spans: • Optimizing streaming speech pipelines (ASR → LLM → TTS) for sub-second response time under production-level load • Instrumenting fine-grained latency profiling (P50/P95) and diagnosing stage-level bottlenecks • Building evaluation frameworks measuring WER stability, partial transcript drift, and real-world inference behavior • Designing failover and liveness mechanisms that improve uptime consistency in distributed AI systems • Fixing silent degradations due to real-world constraints like sample-rate mismatches and back-end variability I’ve contributed to the open-source voice AI framework Pipecat, including: • Fixing Smart Turn v3 prediction failures at non-16 kHz sample rates via high-fidelity resampling (soxr VHQ), improving endpoint detection accuracy from 59% to 94% on telep
Founding AI Engineer, Voice AI
Residency
Forward Deployed Engineer
AI Developer
Masters, Business
Bachelors, Software
