Scientific ML Engineer (Residual Dynamics / Physics-Informed Learning)

About the Role

Our platform is built around a hard rule: don’t learn on lies. We first co-time multiple streams, apply validity gates on each window, cancel what is truly shared (the nuisance), and only then treat what remains as an observable tied to geometry and medium.

This role owns the ML layer after that deterministic contract stack. You will build models that learn residual dynamics—how the “real thing” evolves once timing drift and shared junk have been removed. We want these models to be structured and interpretable (not a black box), so physics-informed and dynamics-learning approaches are on the table.

This is not a “throw a transformer at raw telemetry” role. It’s scientific ML with receipts: train only on valid windows, respect abstentions, and surface explanations that operators can act on.

What You’ll Own

• Residual-state modeling: define a state representation derived from post-cancellation residuals (spatial modes, gradient summaries, spectral components, zone deltas, etc.).
• Structured dynamics learners: implement and evaluate physics-informed / dynamics-learning models (and pragmatic baselines) for forecasting and anomaly detection.
• Trust-aware training: integrate validity scores and window diagnostics into training and evaluation (hard filtering + soft weighting + abstain handling).
• Explainability primitives: produce operator-usable explanations with evidence (“regime shift,” “model violation,” “insufficient validity uptime,” etc.).
• Receipts for ML: ensure every model artifact is auditable: training data hashes, gate criteria, versions, hyperparameters, evaluation suite, and drift monitoring.

What You’ll Do

• Define modeling targets that matter operationally: forecasting (early warning), anomaly detection, and regime change detection on residual dynamics.
• Build residual datasets: raw telemetry → co-timed streams → canceled residual → feature/state tensors, while preserving lineage.
• Run model comparisons and ablations with receipts; benchmark against strong baselines (AR/VAR/state-space).
• Make abstention first-class: don’t train on invalid windows by default; don’t output confident predictions when validity is low; expose “insufficient validity uptime” explicitly.
• Integrate inference into the portal: forecasts with confidence, anomaly events with rationale, timelines that show supporting diagnostics.

Concrete Deliverables

• A residual state spec: what the state is, how it’s computed, and what it represents.
• A model suite (v1): baseline predictors + a structured dynamics learner, with clear evaluation.
• An evaluation harness: fixed splits, backtests, metrics (forecast error, lead time, anomaly precision/recall), and “don’t learn on invalid windows” checks.
• A model receipt system: dataset lineage, gate definitions, version pinning, reproducible training runs.
• A pilot-ready output: one pilot where ML adds concrete value (earlier warnings, fewer false alarms, clearer regime-change explanations).

Required Qualifications

• Strong applied ML experience, including time-series modeling and uncertainty-aware evaluation.
• Experience with dynamical systems learning or physics-informed modeling (e.g., structured dynamics models, neural ODEs, state-space modeling, energy-based constraints).
• Ability to turn messy real signals into stable representations (feature engineering + representation discipline).
• Solid software engineering: reproducible training pipelines, tests, experiment tracking, clean interfaces to backend/portal.

Preferred Qualifications

• Familiarity with signal processing/estimation concepts (coherence, drift, windowing/leakage) enough to respect upstream contracts.
• Experience with anomaly detection for operational systems (alerts, thresholds, fatigue reduction, postmortems).
• Experience with model governance in production: drift detection, retraining criteria, audit artifacts.
• Comfort working across teams (DSP, systems, validation, product) and translating between math and operator language.

How You’ll Be Measured (First 60–90 Days)

• You define and ship a usable residual state and a baseline forecasting/anomaly system that respects validity gating.
• You demonstrate value in one pilot scenario (earlier warning lead time, fewer false positives, or clearer regime-change explanations).
• You integrate training and evaluation into a reproducible pipeline with receipts (someone else can rerun it and get the same result).
• Outputs are explainable: operators can understand what changed and why the model is confident (or abstaining).

Working Style

• You don’t train on garbage and hope it averages out.
• You prefer structured models when structure buys stability and interpretability.
• You treat confidence as a product feature—calibrated, evidence-backed, auditable.

Title & Level

Scientific ML Engineer (Residual Dynamics / Physics-Informed Learning) (senior IC; scope can scale with experience), partnering with DSP/estimation, validation, backend/data, and product/UI.