Beltic verifies agents in real time, on the request path. That means the models behind it run under the same latency budget as the rest of production

• there is no offline batch job to hide behind, and no second chance if a call is slow or wrong.

You can work remotely from anywhere in the Americas, between UTC-8 and UTC-3, or on-site from our San Francisco office - whichever you prefer. We are not hiring outside that band for this role.

What we are looking for

Must-haves

• Years of production ML - building, evaluating, and deploying models, not just research or notebooks
• Strong Python and a modern ML framework (PyTorch or TensorFlow); solid SQL
• Production ML systems end to end: feature pipelines, training, evaluation, and low-latency, real-time serving
• Comfort with low-label and cold-start problems - anomaly detection, unsupervised methods, weak supervision, and building eval sets and labels from scratch
• Data analysis, statistics, and experiment design
• Strong software engineering - you ship versioned, testable, reproducible code, not throwaway models

Strongly preferred

• Applied ML in fraud, risk, abuse, account integrity, or security
• Graph ML / GNNs, sequence or behavioral modeling, or anomaly detection
• LLM / NLP work, and familiarity with AI agents - tool use, function calling, MCP
• Adversarial settings, where the threat evolves against the model
• Explainable, auditable, reproducible ML, with real model versioning

Bonus

• Payment protocols (x402), agent identity (DID / VC), or crypto commitments (Merkle anchoring)

How we hire

1. Intro call (30 minutes)
2. Technical conversation about work you have done (60 minutes)
3. Paid take-home or a pairing session, your choice
4. Final conversation with the founders

We reply to every application.