This is a role for someone who trains models. Not someone who calls them.
We are building classification and detection systems from scratch: fine-tuned transformers, entity detection over structured and unstructured content, and behavioral scoring against a taxonomy we defined ourselves. If your recent work is retrieval-augmented generation, prompt engineering, agent orchestration, or integrating a foundation model API, this is a different discipline and we would be wasting your time.
If you have owned a classifier in production, from data through training through evaluation through the retraining loop, please consider applying.
About SmartVerify
SmartVerify is building the data egress control plane for enterprise AI. We sit inline between AI agents and enterprise data, inspecting every query, enforcing policy in real time, and producing an immutable audit trail.
This is a greenfield build against an existing spec. You have freedom to update the spec as you come in an evaluate the goals. You would be the only ML engineer on staff. You would report directly to the founder, who has a background in this space, and you would have a co-op available to own the evaluation harness and pipeline QA under your direction.
What You Would Own
• The behavioral classification model: multi-class classification of AI agent query intent against our internal taxonomy, producing labels, confidence, and supporting evidence rather than a bare score
• PII and PHI detection over query content and returned data, including span-level identification suitable for audit evidence
• Behavioral risk scoring, combining deterministic request signals with model-derived signals
• Model training, serving, and versioning on SageMaker within the asynchronous inspection path
• The evaluation harness, drift detection, and retraining loop, with a co-op supporting the harness work
• The classification output contract that downstream audit, enrichment, and dashboard consumers depend on
What We Are Looking For
Must-haves are genuinely required. Nice-to-haves are things we expect a strong candidate to pick up here
Must have
Production ML ownership: models you trained, deployed, monitored, and retrained in a live system
Transformer fine-tuning for text classification (BERT family, DistilBERT, or equivalent)
Sequence labelling or named entity recognition for structured entity detection
Python, PyTorch, Hugging Face Transformers
Evaluation rigour: you can explain how you chose thresholds and what you traded away
Comfort working from a written design spec rather than waiting for direction
Authorised to work in Canada or the United States, located in British Columbia or the Seattle area
Nice to have, or will ramp up
Bootstrapping labels through weak supervision, LLM-assisted labelling, or active learning
Anomaly or behavioural detection with sparse or absent labels
SageMaker training jobs and inference endpoints
Regulated domain experience: HIPAA, PCI DSS, GDPR, or SOC 2
Fraud, abuse, or security detection background
Distillation or quantisation to hold an inference latency budget
Kinesis, Kafka, or other streaming pipelines
SQL and query structure parsing
Mentoring a junior engineer or co-op
Who Does Well Here
• You have shipped a model that other systems depended on, and you remember what broke
• You are honest about model limitations rather than defensive, because our customers are auditors and regulators
• You can read an architecture document, disagree with part of it, and say so with a reason
• You are comfortable being the only person in the company who understands this layer, and you document accordingly
• You want ownership more than you want a large team
The Stack
Intelligence layer: PyTorch, Hugging Face Transformers, SageMaker training and inference, evaluation and drift tooling
Infrastructure: AWS, EKS, Terraform, Helm, Prometheus, Grafana
Location and Compensation
• Seattle area or British Columbia. Remote within those regions, with periodic in-person time with the team
• Existing authorization to work in United States is required. We can sponsor work visa and PR in Canada
• Below-market base plus meaningful early-stage equity. We discuss specific numbers early in the process rather than making you guess, and we will not ask you to name a figure first
.png){kind=logo}
