Yuval Herman

Yuval
Herman

Senior Data Scientist, Tel Aviv, Israel

Machine learning for lending. Formerly a CPA.

Email me Read my CV

Machine learning

Models that decide who gets approved, flagged or forecast.

Classification

  • Three Markets

    Credit-risk scoring for lending in three countries, owned from the first feature to production monitoring.

    +0.07 ROC-AUC over baseline

    • LightGBM
    • SHAP
    • Optuna

  • Quiet Signals

    Fraud-detection rules for a lending population.

    Fraud down 2.5%

  • Network Effects

    Behavioral and graph-based features feeding the risk models.

    • Graph features
    • GNNs
    • Neptune

Regression

  • Revaluation

    Bond-risk research combining yield curves, convex optimization and machine learning.

    • Convex optimization
    • Fixed income

Time series

  • Forward Look

    Forecasts for Finance, built on trend and seasonality.

    • Prophet

Causal inference

Not who will pay, but who will pay because we acted.

Uplift modeling

  • Second Call

    A decisioning layer for Collections that picks each customer’s treatment by their estimated response.

    • Uplift models
    • Causal inference

Experimentation

  • A/B testing
  • Imbalanced data
  • Explainability
  • Drift monitoring

GenAI

LLMs that do real work, measured before they ship.

LLM evaluations

  • Seven Readers

    Seven LLMs from the leading AI labs, including Meta, Amazon, Google and Anthropic, put head to head on parsing financial documents, scored with metrics designed for the task.

    Accuracy, price and latency compared

    • Evaluation design
    • Document parsing

Agentic workflows

  • Autopilot

    Agents that train models, validate feature accuracy and parse documents, with a human signing off at the checkpoints.

    • Agent skills
    • MCP
    • AWS Bedrock

Apps

Products I build and ship on my own.

Mobile

  • The Global Min

    A data-science interview-prep app with 900+ adaptive scenario questions across nine topics, written with an LLM content pipeline and automated quality checks.

    On the App Store

    • SwiftUI
    • LLM content pipeline

Tools

What I reach for every day.

Languages

  • Python
  • SQL
  • Gremlin

Building with AI

  • Claude Code
  • MCP
  • AWS Bedrock
  • OpenAI
  • RAG

Modeling

  • LightGBM
  • XGBoost
  • CatBoost
  • scikit-learn
  • PyTorch Geometric
  • Prophet
  • Optuna
  • SHAP

Production

  • MLflow
  • SageMaker
  • Docker
  • Lambda
  • Step Functions
  • Git

Data

  • AWS
  • Snowflake
  • Redshift
  • Neptune

Where would you set the cutoff?

Price your mistakes, then find the threshold that costs least.

Your economics
0.50
Low risk Total cost by cutoff High risk

Your cutoff costs $62,500. The best cutoff, 0.42, costs $55,750.

Defaulters stopped
Good customers declined
Approved
68%
Defaulters stopped
61%
Good customers declined
26%

Synthetic book of 496 applicants, model AUC 0.78.

Experience

  1. 2020–Present Data Scientist, Fido Credit Solutions
  2. 2018–2020 Software Developer & Blockchain Administrator, Dov & Co, CPA
  3. 2015–2018 Financial Auditor, UHY Israel

See the full CV

Writing

  1. Reading an ROC curve as a lending decision

    Every point on a credit model's ROC curve is a policy someone has to sign off on. Here is how to read it that way.

All writing