Senior Machine Learning Engineer
Summary
Senior ML Engineer at a pre-seed stealth startup building privacy-preserving, on-device voice biometric authentication. Day to day: run high-velocity experiments on voice embedding models (speaker verification, liveness, age estimation), compress models via quantization/pruning/distillation to fit <5MB on-device, build ML infra, and ship to iOS/Android with CoreML/TFLite/ONNX using PyTorch/JAX.
Senior ML Engineer
Be the ML force multiplier at a funded stealth startup solving hard on-device AI problems.
ABOUT THE COMPANY
A stealth startup in the identity space building privacy-preserving, on-device authentication via voice biometrics. Backed by $2.4M in pre-seed funding with large enterprise clients already in the pipeline. The core technical challenge: compressing 5GB ML models to under 5MB without sacrificing accuracy. Small team, real traction, massive upside.
WHY THIS ROLE EXISTS
We need someone who brings a different gear — modern tooling, fast experimentation, and a bias toward shipping. You won't be inheriting a broken system; you'll be raising the ceiling on what this team can do. This is the kind of role where the right person becomes indispensable within 90 days.
WHAT YOU'LL OWN
Drive high-velocity experimentation on voice embedding models — speaker verification, liveness detection, age estimation
Own model compression end-to-end — quantization, pruning, distillation — to hit aggressive on-device size and latency constraints
Build modern ML infrastructure — experiment tracking, fast eval loops, reproducible pipelines — so the team can iterate at a different pace
Deploy optimized models to iOS and Android via CoreML, TFLite, or ONNX Runtime and own their performance in production
Partner with the co-founder to define the ML roadmap and unblock enterprise client integrations
YOU ARE THE RIGHT FIT IF
You've shipped production ML models and measure yourself by what's deployed, not what's in a notebook
You run 10 experiments where others run 2 — fast iteration is your default, not a mode you shift into
You live in modern tooling — Weights & Biases, HuggingFace, modal, or equivalent — and have strong opinions about what a good ML stack looks like
5+ years of applied ML engineering with deep PyTorch or JAX fluency across the full training loop
Hands-on experience with model compression — you've hit real size and latency walls and engineered your way through them
Startup-ready: you set your own direction, don't wait to be unblocked, and thrive when the brief is "figure it out"
NICE TO HAVE
Background in audio/speech ML — speaker verification, voice activity detection, or audio embeddings
Experience with on-device deployment via CoreML, TFLite, or ONNX
Familiarity with privacy-preserving ML — federated learning, on-device inference, differential privacy
Prior experience at an early-stage startup — you know the difference between building for scale and building to learn
