• Binary classifiers under heavy class
imbalance. Fraud base rates are low. You will build GBMs (XGBoost,
LightGBM) and shallow neural models that have to perform at the low-FPR
operating points clients actually care about — not at default thresholds.
• Feature engineering across signal families. Device,
behavioral, location, image, telecom and alternate-data signals, plus velocity
aggregations across multiple time windows. You will learn to turn raw telemetry
into features that survive adversarial drift.
• Anomaly and unsupervised detection. Isolation
Forest, auto encoders, PCA residuals, peer-group outliers — for the cold-start
cases where labeled fraud is weeks away.
• Graph features (growing scope).Contribute
to our graph-based fraud and community-detection work. You will start with
graph-derived tabular features and can grow deeper into graph modelling if the
work suits you.
• Deployment. Package models for our
real-time decisions path. Understand latency budgets and feature availability
at scoring time (no leakage, no missing features in production).
• Monitoring. PSI, CSI, score-distribution
drift, feature drift, fraud-capture curves by client. Own the weekly
model-health review for at least one deployed model within six months.
• Threshold economics. Work with clients to
translate fraud capture ↔ customer-friction trade-offs into operating points
they can sign off on. This is where most fraud models actually fail, and where
you will learn the most.
• With engineering, on feature pipelines,
offline/online parity and real-time inference.
• With product and clients, on what a model
is actually allowed to flag, block or auto-approve. You will sit in review
calls with bank fraud heads.
• 1–3 years of hands-on ML experience building
supervised models on real data (not only coursework or Kaggle).
• Strong Python and SQL. Comfortable with
pandas/Polar, scikit-learn, XGBoost or LightGBM, and writing non-trivial SQL
against large tables.
• Statistical intuition. You know why
accuracy is the wrong metric for fraud, can explain ROC vs PR curves, and
understand calibration and threshold selection.
• Engineering hygiene. Git, code review,
reproducible experiments, basic CI. Your models should run when someone else
checks out your branch.
• Clear writing. You can write a one-page
memo that a non-technical product manager and a fraud-ops lead both walk away
understanding.
• Exposure to fraud, risk, credit, AML or payments
data — at a bank, FinTech, card network or risk-tech vendor.
• Experience with imbalanced-class techniques
beyond naive resampling (focal loss, cost-sensitive learning, threshold-moving,
calibration).
• Experience with real-time feature serving, model
monitoring or MLOps tooling (MLflow, Feast, SageMaker, Vertex, or equivalents).
• Exposure to graph data, NLP/NER, or geospatial
features.
• A public repo, paper, competition placement or
blog post that shows how you think.
• Compensation benchmark to top-quartile Indian
FinTech; ESOPs for all full-time employees with a clear vesting schedule.
• Health insurance for you and your family, and a
meaningful learning & conference budget.
• Access to real, messy, high-volume Indian
financial data and the compute to do something interesting with it.
Skills Required
- 1-3 years of hands-on ML experience building supervised models on real data
- Strong Python
- Strong SQL and writing non-trivial queries against large tables
- Experience with pandas or Polars
- Experience with scikit-learn, XGBoost or LightGBM
- Statistical intuition (ROC vs PR, calibration, threshold selection)
- Engineering hygiene: Git, code review, reproducible experiments, basic CI
- Clear technical writing for non-technical stakeholders
- Exposure to fraud, risk, credit, AML or payments data
- Experience with imbalanced-class techniques beyond naive resampling (focal loss, cost-sensitive learning, threshold-moving, calibration)
- Experience with real-time feature serving, model monitoring or MLOps tooling (MLflow, Feast, SageMaker, Vertex, or equivalents)
- Exposure to graph data, NLP/NER, or geospatial features
- Public repo, paper, competition placement or blog post demonstrating work
What We Do
Cedar is a global strategy consulting, research, and analytics firm with a 35-year track record and clients across multiple industry sectors. Since 1985, its teams have assisted clients in strategy, process innovation, strategic human capital, and business technology, with a strong focus on the Financial Services sector. As a full-suite management consulting firm, Cedar assists clients from strategy formulation to execution and implementation.
