Data Scientist

The Role

You own AI systems end to end. From the speech-to-text models that turn audio into text, to the diarization that separates and identifies speakers, to the agentic layer that turns conversation into memory and action, to the observability and evaluation that keep all of it honest in production. This is a wide role by design. You will own model selection, serving, and production reliability. If you want to tune one model and ignore the system around it, this is not the role.

What You Will Own

•     Speech-to-text. Evaluate, integrate, and optimize STT models across cloud and self-hosted. Drive accuracy and cost trade-offs with ground-truth metrics.

•     Speaker diarization and identification. Push accuracy on hard, real-world, multi-speaker audio.

•     Agentic AI. Build the memory and retrieval pipeline, LLM orchestration, and the agent workflows that sit on top of captured conversation.

•     Model serving and infrastructure. Stand up and optimize self-hosted serving (vLLM, Triton class). Own latency, throughput, and cost per user.

Observability

An always-on wearable means models run in production every second, on messy real-world audio. You own the visibility into that.

•     Instrument the full audio-to-memory pipeline: STT, diarization, retrieval, and LLM calls.

•     Define and track model-quality SLOs in production: transcription drift, diarization error over time, retrieval relevance, latency, throughput, and cost per user.

•     Build dashboards and alerting so model degradation is caught before users feel it.

•     Trace failures across a distributed, always-on system using metrics, logs, and traces.

•     Close the loop. Production signals feed back into evaluation and model selection.

Evaluation

We do not ship what we cannot measure. You own the systems that prove a model is actually better, not just newer.

•     Build and own ground-truth evaluation harnesses for every model in the stack.

•     Measure with real metrics: WER for transcription, DER for diarization, Recall and F1 for retrieval and speaker identification.

•     Build and maintain labeled benchmark datasets that reflect real, messy, multi-speaker audio.

•     Run regression and A/B evaluations on every model swap, prompt change, or pipeline update. Nothing ships on a vibe.

•     Reject anecdotal proxies, single confidence scores, and cherry-picked examples as evidence of quality.

What We Are Looking For

•     3 to 5 years as an AI/ML engineer with production systems behind you. Engineering and production experience is non-negotiable.

•     Depth across the modern AI stack: LLMs, speech models, vector retrieval, model serving.

•     Strong software engineering. You write code that ships and survives contact with real users.

•     Fluency in Python and the production ML ecosystem.

•     Comfort with cloud infrastructure (GCP a plus) and containerized deployment on Kubernetes.

•     A working command of observability and evaluation. You measure first and trust metrics over intuition.

•     First-principles reasoning and metric discipline.

Nice to Have

•     Research background or publications. A strong signal, not a substitute for production work.

•     Audio and speech ML experience (STT, diarization, voice).

•     Experience self-hosting and optimizing open models.

•     Experience with LLM gateway and agent orchestration patterns.

•     Experience building eval harnesses or production model-monitoring systems.

Requirements

Agentic work is must. Audio is good to have

. Self hosting models is a must

 Experience with LLM gateway and agent orchestration is a must have