SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Sunset is a rapidly scaling company (multi-eight-figure run rate in months) that partners with frontier AI labs by providing de-identified enterprise data as training data. The company transforms sensitive internal business data across multiple modalities—messages, documents, tables, files, images, metadata—into valuable datasets while preserving structure and meaning.
As a Machine Learning Engineer, you will own and improve core components of the de-identification pipeline, including named-entity recognition, entity and identity resolution, structured extraction, classification, and semantic review. This is an applied, production-facing role where you'll study real errors, form hypotheses, build datasets and experiments, improve or replace models, and ship results into a live pipeline.
Key responsibilities include:
- Owning specific de-identification systems and transforming model failures into prioritized improvement roadmaps
- Designing active-learning loops combining model sweeps, LLM-assisted review, clustering, and uncertainty signals
- Building representative datasets, benchmarks, and decision-relevant metrics to reveal strengths, weaknesses, and failure costs
- Selecting and combining deterministic rules, classical ML, fine-tuning, embeddings, multimodal models, and LLM approaches based on evidence
- Designing rigorous experiments, analyzing precision-recall tradeoffs, and explaining which changes are real and generalizable
- Productionizing improvements with reproducible artifacts, evaluation evidence, instrumentation, and safe rollout
- Optimizing inference cost, latency, and throughput without hiding quality regressions
- Building high-fidelity evaluation environments with seeded failure modes and programmatic verifiers
- Partnering with Applied Science, Data/Product Engineering, and Security teams on measurement and risk
- Using AI engineering tools fluently while verifying their output
Success means model improvements that generalize beyond development examples, credible gains in precision/recall/F1 across priority modalities, reduced high-risk misses without over-redaction, and faster, more repeatable paths from error discovery to production improvements.
You should have 3+ years of professional ML or software engineering experience, including improving models in production. Startup experience and comfort with evolving requirements are valued. You're a strong Python engineer who works inside data pipelines and production systems, not just notebooks. You understand precision, recall, F1, calibration, thresholding, class imbalance, imperfect labels, distribution shift, and representative evaluation deeply. You use modern AI tools fluently and verify their output.
