Lead Applied Scientist – NLP and Ranking

Department/Project Description

GlobalLogic takes an opportunity to partner with a leading global threat intelligence platform to design and deliver an enterprise-scale, Human-in-the-Loop (HITL) AI agent platform. The initiative centers on accelerating expert analyst decision-making by augmenting high-complexity threat intelligence workflows with cutting-edge generative and agentic AI systems.

Job Description

Essential qualifications

MSc or PhD in NLP, information retrieval or machine learning with publications, or equivalent industry research experience.

5+ years of applied ML, with models in production use.

Deep experience fine-tuning transformer encoders with PyTorch and Hugging Face.

Multi-label and imbalanced classification, probability calibration and threshold optimization.

Learning from user feedback: learning-to-rank, active learning, bandit/exploration methods or counterfactual evaluation.

Rigorous experiment design and statistical analysis.

Desirable qualifications

ModernBERT or other long-context encoder models.

Multi-stage ranking or recommender systems.

LLM post-training (SFT, LoRA, DPO), reward or judge models, and reranker or embedding fine-tuning.

Vertex AI Training and Pipelines.

Experience in security, threat intelligence or fraud.

Job Responsibilities

Key responsibilities

Fine-tune, evaluate and calibrate ModernBERT for multi-label alert classification and high-recall shortlisting.

Design the two-stage triage cascade: model thresholds, the hand-off to the LLM, and how each stage is measured.

Convert analyst feedback into training data through label cleaning, active learning and defined retraining criteria.

Correct for selection bias through exploration sampling of rejected alerts and counterfactual evaluation.

Design the dynamic few-shot strategy:

how decision points are represented

how examples are selected

how conflicting or stale human decisions are handled.

Define the feedback and reward schema with the Lead Agent Engineer.

Set offline and online evaluation, and review retrained models before production promotion.