Hari Gebre

Sheet 01 · Boston, MA

Hari Gebre

Data ScientistMachine Learning & Generative AI

Data scientist with 8+ years of experience building predictive models and analytical solutions in Python, SQL and PySpark for financial services and customer-behavior problems. Works across Scikit-learn, XGBoost, TensorFlow, PyTorch and LLM tooling, and turns model results into clear business insight.

Model pipeline Records move through features and an XGBoost model, tracked in MLflow and reported in Power BI, with PySpark and SQL along the way. pipeline records features model PySpark · SQL Pandas · NumPy XGBoost MLflow Power BI tracked, reproducible, reported
8+ yrsin data science and analytics
500K+financial records processed with PySpark
250K+banking transactions modeled with XGBoost
200K+consumer accounts analyzed at Synchrony

02 · About

The work so far

Risk, segmentation, generative AI

I build models that show a bank where its risk sits, and explain them in terms a business team can act on.

At PNC I build XGBoost models for risk assessment and keep the work reproducible by tracking experiments in MLflow. At Synchrony I segmented consumer credit customers with K-Means and forecast transaction volume with ARIMA.

I am also working with generative AI and LLM tooling, including RAG, embeddings, semantic search, LangChain and LangGraph. The sketch shows the retrieval pattern they share.

Retrieval flow Documents become embeddings, a retrieval step selects passages, and a language model writes the answer; a dashed LangGraph loop returns from the answer to the documents. documents embeddings retrieval LLM answer LangGraph loop
Retrieval pattern: RAG with embeddings and semantic search

03 · Experience

Roles

2018 to present

  1. PNC logo

    Data Scientist

    PNC · Boston, MA
    • Built XGBoost models in Python to flag unusual transaction behavior and customer activity patterns for risk assessment.
    • Prepared large financial datasets with PySpark: cleansing, transformation and aggregation for machine learning.
    • Engineered behavioral features with SQL and selected variables from 15+ candidate attributes.
    • Built Power BI reporting on 12+ financial and risk KPIs, and presented findings in monthly risk reviews.
    • Python
    • SQL
    • XGBoost
    • PySpark
    • MLflow
    • Power BI
  2. Synchrony logo

    Data Scientist

    Synchrony · USA
    • Analyzed consumer account records with Python and SQL to find customer behavior patterns across portfolios.
    • Applied K-Means clustering to 10+ behavioral features and developed 4 customer segments.
    • Built ARIMA forecasts on 18+ months of history to track transaction-volume and customer-activity trends.
    • Ran hypothesis tests and A/B analysis across 2 customer groups, and built Tableau dashboards on 10+ portfolio KPIs.
    • Python
    • SQL
    • K-Means
    • Logistic Regression
    • ARIMA
    • Tableau
  3. DICEUS

    Data Analyst

    DICEUS · USA
    • Reviewed and standardized client records, identifying duplicate, incomplete and inconsistent data.
    • Created reporting views covering 4+ business KPIs for senior team members.
    • Validated client information across sources and prepared analysis-ready datasets for reporting.
  4. Allstate logo

    Business Analyst

    Allstate · Lowell, MA
    • Used Excel and SQL to review policy and customer records for missing information, duplicates and inconsistencies.
    • Prepared recurring reports on policy activity, customer requests and operational KPIs.
    • Supported UAT with test cases and defect documentation, and maintained BRD/FRD documents and requirement trackers.
    • Excel
    • SQL

Education

Metropolitan State University, CO

Computer Information System

04 · Projects

Selected work

Four full sheets, six more below

persuadable sure thing lost cause sleeping dog contact effect → churn risk

Sheet 01 · Causal ML

upliftpolicy

Picks customers to contact by how much a contact changes their decision, not by how likely they are to leave. It also flags the customers a contact would drive away.

  • Top 30% by uplift adds 4.47 points of retention; top 30% by churn score loses 0.23 points, on simulated customers.
  • Sleeping dogs make up 37% of the churn-ranked contact list but 4% of the uplift-ranked one.
  • Python
  • Uplift modeling
  • T-learner
  • Doubly-robust estimation
invoice_total policy_number incident_date damage_description threshold = review cutoff

Sheet 02 · Document AI

fieldextract

Sends each extracted field to automatic approval or human review, with the cutoff set by the cost of an error, so operations can see which fields need a person.

  • Aiming for a 1% field-error rate, it auto-approves 43.8% of fields, with a realised error of 0.87% on held-out data.
  • Per-field routing costs 1.78x less than one tuned global threshold on the same 4,817 reviews.
  • Python
  • Isotonic calibration
  • Cost-based routing
  • PII redaction
national regions sub-regions parts sum to the whole coherent by design

Sheet 03 · Forecasting

hierforecast

Energy demand forecasts that add up: regional and sub-regional forecasts sum exactly to the national figure, and the upper tail is scored on its own.

  • On the p95 pinball loss, reconciled forecasts score 14.3% lower than bottom-up.
  • Independent forecasts disagreed by up to 612 MW; after reconciliation the gap is round-off.
  • Python
  • Reconciliation
  • Quantile forecasting
  • Pinball loss
direct overrideindirect retrievalroleplay framingencoding obfuscationtool-output poisoning

Sheet 04 · LLM security

injectbench

Measures how often prompt-injection attacks get through, class by class, and which guardrail layers actually reduce that rate.

  • Attacks get through 10.8% of the time with the full guardrail stack, against 87.6% with none.
  • Roleplay framing is the weakest class at 46.5%, a figure a single overall rate would hide.
  • Python
  • Ablation study
  • Guardrails
  • Prompt injection

More sheets

  • Sheet 05 · Vector search · Rust

    hnswops

    A vector-search index written from scratch in Rust, with memory-mapped snapshots, online recall monitoring and zero-downtime re-embedding.

    Source
  • Sheet 06 · Security analytics · Rust

    huntql

    A pipe-based threat-hunting query language for Zeek logs, with a typed planner, a streaming executor and beaconing detection.

    Source
  • Sheet 07 · Binary analysis · Rust

    binscope

    Static triage for ELF, Mach-O and PE binaries: sections, entropy, imports, capability rules and ssdeep and TLSH hashes.

    Source
  • Sheet 08 · Security tooling · Python

    fwaudit

    Offline firmware auditor for embedded Linux and IoT. It unpacks images in memory and reports hardcoded credentials and weak services.

    Source
  • Sheet 09 · Program analysis · Python

    solsentry

    A static security analyzer for Solidity that reports eight smart-contract bug classes as text, JSON or SARIF.

    Source
  • Sheet 10 · Go

    sysprobe

    An eBPF profiler for Linux that records per-process syscall, run-queue and TCP connect latency as Prometheus histograms.

    Source

05 · Toolkit

Skills

Strongest tools by group

Programming & Analysis

  • Python
  • SQL
  • Pandas
  • NumPy
  • PySpark
  • Statistical Analysis
  • A/B Testing

Machine Learning

  • Scikit-learn
  • XGBoost
  • CatBoost
  • Logistic Regression
  • K-Means
  • Feature Engineering
  • Time-Series Forecasting

Deep Learning & NLP

  • TensorFlow
  • PyTorch
  • Keras
  • Transformers
  • Hugging Face
  • CNN
  • RNN

Generative AI & LLM

  • RAG
  • Embeddings
  • Semantic Search
  • LangChain
  • LangGraph
  • Pinecone
  • LLM Evaluation

Big Data & Engineering

  • Apache Airflow
  • Kafka
  • ETL/ELT
  • dbt
  • Data Warehousing
  • Snowflake

Cloud & Deployment

  • AWS Lambda
  • AWS EC2
  • Docker
  • Kubernetes
  • CI/CD
  • MLflow

Analytics & Visualization

  • Power BI
  • Tableau
  • Excel
  • Power Query
  • Plotly
  • KPI Reporting
  • Data Validation

Databases

  • PostgreSQL
  • Microsoft SQL Server
  • MySQL
  • Oracle

06 · Contact

Let's talk

I'm happy to talk about machine learning, predictive modeling and AI engineering roles. Email is the quickest way to reach me.

gebrehari01@gmail.com