Machine Learning Engineer - Speech & Natural Language

Primordial Labs•New Haven, CT

32d•$150,000 - $215,000•Remote

About The Position

Anura turns a warfighter's spoken intent into robotic action, and it does it on hardware carried into the field, with no cloud connection and no second chances. You own the models that make natural language work at the tactical edge: speech recognition, natural language understanding, and the full path from candidate approach to a trained, hardened model running reliably on constrained compute. This is not a foundational research role. We are building the most accurate and lowest-latency language stack that will fit in a soldier's ruck, and we're looking for someone who has already done work like this and can show it.

Requirements

Proven, Not Learning on the Job: You've shipped ML into production and can walk through exactly what you built, what broke, and what you measured. This role needs real ownership from day one.
Deployment-Minded: A model that only runs in a notebook isn't finished. You think about memory ceilings, inference paths, and startup time as naturally as you think about loss curves.
Data-Obsessed: You know that improving the data usually beats tuning the architecture, and when the data you need doesn't exist, your instinct is to build the tooling to create it rather than wait on someone else.
Language and Speech Native: Speech recognition and language understanding are your home turf. You have opinions about decoding strategies and about the ways standard accuracy metrics mislead you.
Pragmatist Over Purist: You reach for the smallest thing that solves the problem, and you can tell the difference between a model that benchmarks well and one that works for a user under pressure.
Remote-Ready: You communicate clearly in writing, work independently across time zones, and keep the team synced without being asked.
Track Record: Substantial hands-on experience training, optimizing, and deploying ML models in production systems. We're flexible on years and title, biased toward senior, and inflexible on demonstrated results.
Speech and Language Depth: Deep, hands-on experience across both speech recognition and natural language understanding, including adapting models to a specialized domain.
Core Skills: Deep proficiency in Python and modern deep learning frameworks, including custom training loops, data pipelines, and evaluation harnesses.
Optimization: Demonstrated experience making models meaningfully faster or smaller, with the benchmarks to back it up.
Productionization: Hands-on work moving models out of framework and research code into an efficient, deployable inference path.
Education: Degree in CS, EE, Computational Linguistics, or a related technical field, or equivalent proof you can do the work.
Applicants must have current authorization to work in the U.S., and we are not able to provide visa sponsorship for this role.

Nice To Haves

Deploying models on embedded or resource-constrained GPU hardware
Defense, robotics, or autonomy experience
Model lifecycle tooling: experiment tracking, model registries, automated evaluation

Responsibilities

Evaluate and Decide: Assess candidate architectures and modeling approaches against the criteria that actually matter operationally: accuracy under degraded conditions, latency, and footprint. Make the call and defend it with data, not vibes.
Build the Data Engine: Develop synthetic data generation tooling and augmentation pipelines, and curate the data we collect from the field. Our accuracy advantage comes from data we build ourselves, and you'll own that machinery.
Train for the Domain: Train and fine-tune models that understand how operators actually talk, in noisy environments and with degraded audio, using language no general-purpose model has seen enough of.
Optimize Ruthlessly: Every millisecond of latency and every megabyte of memory is a design decision with operational consequences. You'll find the wins and prove them with benchmarks.
Mature the Deployment: Turn research-grade code into something we can ship: model weights served directly through an efficient production runtime, with the experimental scaffolding stripped out.
Close the Loop: Own how we measure success. Build evaluation that reflects whether an operator's command actually worked, not just how a model scores on a test set, and turn field failures into test cases, training data, and fixes that stick.
Sharpen Our MLOps: Improve how we train, track, evaluate, and release models. At minimum, be relentlessly vocal about what's broken.