Founding ML and Evaluation Engineer

QLUU Lab Inc.; Berkeley (CA); Remote; USD 120,000 - 180,000; Full time

Job summary

Qluu Lab, Inc. is seeking a founding ML and Evaluation Engineer to own the critical question: does the model actually work, and can we prove it?

You will be the independent check on our code and drive improvements, with remote-first collaboration and field tests in the U.S. In the first 90 days you will build an independent evaluation harness, benchmark against baselines, train and ablate on GPUs, export to ONNX for edge devices, and own the human-review protocol.

Qualifications

• Trained sequence or trajectory models in PyTorch and can read a loss curve and tell when it is lying.
• Understanding of grouped splits, leakage, bootstrap confidence intervals, multiple-comparison control.
• Write reproducible code with pinned environments, hashed data, deterministic seeds where it matters.
• Comfort with a small team, ambiguity, and shipping something on real hardware.
• Nice to have: ONNX/TensorRT/ONNX Runtime and edge deployment experience.

Responsibilities

• Own the evaluation framework and build an independent harness separating data loading, transforms, and metrics from training code.
• Measure the deployed selector against strong baselines and publish where Foresight wins or loses by domain.
• Train and ablate with GPUs, ensuring clean provenance for every checkpoint (config, data hashes, splits, seeds).
• Take the model to edge devices with ONNX export, latency/memory measurements on Raspberry Pi 5 and Nvidia Jetson.
• Own the human-review protocol with blinded samples and reviewer agreement.

Skills

Sequence/trajectory models

PyTorch experience

Statistics knowledge

Reproducible code practices

Edge deployment

Tools

PyTorch

ONNX

ONNX Runtime

TensorRT

Job description

Qluu Lab, Inc. | Remote-first, with field tests in the U.S. | Full-time

Founding ML and Evaluation Engineer

About us

Qluu Lab builds Foresight, a prediction model for the motion and behavior of tracked objects: where something is going and what it is likely to do next. It works from track data, runs on a ground machine, and has a small export for edge computers. We are a software company, and this is the first dedicated ML hire.

The job

You will own the question every serious ML team has to answer: does it actually work, and can we prove it? You will be the independent check on our code, and then you will make it better.

In your first 90 days:

• Build the independent evaluation harness. Separate data loading, transforms and metrics from the training code. Pre-register pass/fail gates before looking at test data.
• Measure the deployed selector against strong baselines (constant velocity, Kalman, interacting multiple models) with scenario-clustered confidence intervals. Publish where Foresight wins and where it loses, per domain.
• Train and ablate. Run noise-robustness and selector experiments on GPUs we already have, with clean provenance for every checkpoint (config, data hashes, splits, seeds).
• Take the model to the edge. ONNX export, quantization parity tests, and measured latency and memory on a Raspberry Pi 5 and an Nvidia Jetson. Measured, not estimated.
• Own the human-review protocol: blinded review samples, reviewer agreement, written failure cases.

You might be a fit if you

• Have trained sequence or trajectory models in PyTorch (motion forecasting, time series, tracking, or similar) and can read a loss curve and tell when it is lying.
• Care more about a result that survives a hostile evaluation than about a good-looking number.
• Know your statistics: grouped splits, leakage, bootstrap confidence intervals, multiple-comparison control.
• Write reproducible code: pinned environments, hashed data, deterministic seeds where it matters.
• Are comfortable with a small team, ambiguity, and shipping something on real hardware.

Nice to have: Kalman/IMM tracking, state-space or Mamba-style models, ONNX/TensorRT/ONNX Runtime, mixed-precision training, defense or simulation data (DIS/HLA), prior work with export-controlled data.

Why this role

• You report to the CTO and have real authority over what we are allowed to claim.
• You will see every result, good or bad, before anyone outside the company does.
• Your evaluation work will be what we show investors and customers.
• Early equity and a direct line to the technical decisions.

Requirements and logistics

• U.S. person required. You must be a U.S. citizen, lawful permanent resident, or protected individual.
• Willingness to go through background screening. Ability to obtain a U.S. security clearance is a plus.
• Some travel for field tests.
• Qluu Lab is an equal opportunity employer.