You'll join at an inflection point. We're rebuilding our detection stack on a modern decision-engine architecture — inline hot-path decisioning, a complete decision audit ledger, and entity-level risk state — with active programs in social-engineering, phishing, and brand-impersonation detection.
What You Will Do
Design, train, evaluate, and operate text classifiers end-to-end: data curation, labeling, offline evaluation, deployment, and production drift monitoring.
Run shadow-mode evaluations against labeled fraud corpora and human-reviewed legitimate controls before any model earns enforcement authority; report precision, recall, and calibration per abuse archetype.
Build detection that holds up against adversarial content: obfuscated and confusable brand text, URL and domain evasion, and cross-account template fingerprinting.
Serve detection models inline on the message send path under strict latency budgets, with deliberate fail-open/fail-closed behavior per dependency.
Power human-in-the-loop fraud review: reviewer tooling, label quality, and turning fraud dispositions into replay and training signals.
Collaborate with fraud operations, trust & safety, and cross-functional engineering teams.
Participate in code reviews and help raise code quality and standards.
What You Bring
3+ years of software development and/or machine learning experience.
Strong proficiency in async Python webservice frameworks (e.g. FastAPI).
Production experience building and operating text classification systems in fraud, spam, abuse, or trust & safety.
Rigorous evaluation methodology: labeled corpora, holdout sets, precision/recall tradeoffs, and calibration.
Experience using LLMs as classifiers or judges within a larger decision pipeline.
Experience deploying and maintaining ML models in production.
Excellent communication skills and the ability to work effectively with cross-functional teams.
Nice to Have
Familiarity with A2P messaging abuse: artificially inflated traffic, smishing, brand impersonation, and 10DLC / toll-free verification regimes.
Adversarial ML experience: evasion-resilient detection, obfuscation handling, or attack-surface hardening.
High-scale data experience with ClickHouse, PostgreSQL, Redis, and streaming systems such as Kafka.
Experience with a cloud-native stack (e.g. Docker, Kubernetes, Prometheus).
Experience with distributed batch scoring (e.g. Ray).