About role: You'll be the engineering owner on the speech team, working alongside two ML engineers and building everything around the models — ingestion, training infrastructure, real-time serving, and the integration into our existing platform. This is a backend engineering role. You don't need to train models. You need to build the systems that make trained models useful in production. Responsibilities: • Build the audio data pipeline: mine our call archive, transcode, resample, segment, deduplicate and quality-filter at scale • Build and operate real-time inference services with hard latency targets, including streaming, cancellation and mid-utterance interruption • Integrate speech services into our existing Java-based platform and telephony infrastructure • Instrument the full latency budget end to end and find where the milliseconds go • Stand up training infrastructure — GPU scheduling, checkpointing, experiment tracking, reproducibility • Own compute cost and concurrency economics: how many simultaneous calls per GPU, and how to improve it • Support on-premise deployment for clients who can't send data outside their network Requirements Must haves: • 3–4 years building and operating production backend systems • Strong Java — you've owned services in production, not just contributed to them • Working Python — enough to build data pipelines and integrate with ML tooling • Real-time or low-latency systems experience: streaming APIs, WebSocket or gRPC streaming, concurrency, backpressure • Data pipelines at scale (Airflow, Dagster, Spark or equivalent) • Docker and Kubernetes in production • Cloud infrastructure (AWS/GCP/Azure) • Comfortable debugging performance: profiling, latency percentiles, throughput under load Nice to have: • Serving ML models in production (Triton, vLLM, TorchServe) • Audio tooling — ffmpeg, sox, codecs, resampling • Telephony — SIP, Asterisk/FreeSWITCH, media servers, narrowband codecs • MLOps tooling: MLflow, Weights & Biases, DVC • GPU-aware infrastructure work