Essohanam Josué Kpatcha | Data & AI Engineer

About

Engineer who thinks
like a consultant

I'm a final-year Data Engineering student at ENSA Al Hoceima (Morocco), obsessed with one question: what does this data actually tell us - and what should the business do about it?

Four years translating data into decisions: churn diagnostics for energy utilities, financial health scoring for 9,600 African SMEs, national connectivity demand modeling for Togo. Behind every insight - production pipelines on Azure Databricks, cloud-native platforms, and executive dashboards built to communicate, not just visualize. Two internships in applied AI confirmed that the hardest part is never the model - it's framing the right problem.

What makes me different: I bring engineering precision to consulting questions. I don't just build the analysis - I deliver the recommendation.

Essohanam Josué KPATCHA - Data & AI Engineer

8 Consulting-grade case studies delivered end-to-end

0.86 ROC-AUC - Zindi Africa Togo Internet Uptake Prediction

0.96 Precision - BCG X PowerCo Churn Analysis (XGBoost)

0.82 Weighted F1 - African SME Financial Health Prediction

Experience

Data & AI Engineer Intern - End-of-Studies Project (PFE)

GeTeam

Current

Mar 2026 – Present
Casablanca, Morocco

Developed and deployed an end-to-end AI system for EU Pay Transparency Directive 2023/970 compliance, translating a complex regulatory problem into a production solution used by HR teams.
Fine-tuned GLiNER/XLM-RoBERTa (183M params, PEFT) on 1,100 annotated examples, achieving F1 = 65.7% micro-avg (+15 pts vs. zero-shot) on gender bias detection in French HR documents, no benchmark existed.
Orchestrated 3 parallel LLM agents (Bifrost/OpenRouter) with consensus scoring and full Langfuse observability; applied multiple clustering algorithms (Ward Hierarchical, KMeans, GM) and model selection with IF-THEN rule extraction for interpretable audit reports.
Delivered a Full Stack application (FastAPI + React with SSE streaming and 2D cluster visualizations), Docker images, and GitHub Actions CI/CD pipeline.

Python PyTorch GLiNER LangChain Langfuse Bifrost FastAPI React Docker GitLab JIRA

AI & Data Engineering Intern

ISICOD

Jun 2025 – Sep 2025 · 4 months
Rabat, Morocco

Co-built an end-to-end biometric AI platform for fingerprint classification & ridge counting across 12 classes. Sole implementer of the Human-in-the-Loop retraining system enabling continuous model improvement in production.

PyTorch FastAPI React Computer Vision

Machine Learning Engineer Intern

3D Smart Factory

Jun 2024 – Aug 2024 · 2 months
Mohammedia, Morocco

Applied research on 3D medical image segmentation. Implemented custom KAN layers integrated into a 3D U-Net - achieving +3% segmentation accuracy gain over the standard baseline on MRI datasets.

PyTorch TensorFlow 3D Deep Learning Medical AI

Case Studies

Work that speaks
in business outcomes

Each project structured as a consulting engagement: problem framed, approach rigorous, impact quantified.

★ Featured Case Study BCG X Analytics Simulation

Why PowerCo's Customers Churn - and What To Do About It

Energy utility · 14,600 customers · 193,000 pricing records · One bold hypothesis reframed.

View on GitHub

Problem

PowerCo (energy utility) believed price sensitivity was the primary churn driver. BCG was asked to investigate and recommend a targeted retention strategy.

Approach

Reframed the hypothesis. Analyzed 14,600 customers + 193,000 pricing records. Engineered 50+ features. Benchmarked Random Forest, Gradient Boosting, XGBoost with rigorous CV.

Impact

Price is NOT the main driver - margin and customer segment are. Prioritize retention on low-margin, high-churn segments. XGBoost: ROC-AUC = 0.68, Precision = 0.96.

★ Featured Case Study Advanced Analytics · Strategic Recommendations

NovaMart Retail Intelligence - End-to-End Advanced Analytics

Retail · 504,038 transactions · 800 stores · 316K loyalty members · 36-month gross margin decline diagnosed.

View on GitHub

Problem

A retail group faced a sustained gross margin decline from 42.0% to 31.8% over 36 months across 504,038 transactions, 800 stores, and 316K loyalty members. Root cause was unknown.

Approach

Diagnosed COGS inflation as root cause via OLS regression (R² = 0.917, p < 0.001). Tested 5 pre-specified hypotheses (ANOVA, Welch t-test, Mann-Whitney U, Pearson r) with effect sizes. Built 4 production ML models: churn XGBoost, store RF, Marketing Mix Model with Adstock, price elasticity OLS. SHAP + MLflow throughout.

Impact

8 RICE-scored recommendations totaling $32.4M addressable annual impact: $23.4M urban digital reallocation, Silver tier loyalty rescue ($8-12M at risk). Churn AUC = 0.98, store RF R² = 0.87. Streamlit executive dashboard with GenAI copilot (Claude API, tool-use pattern).

BCG GenAI Consulting Simulation

AI Financial Analyst for Global Finance Corp.

Problem

GFC needed to analyze hundreds of 10-K filings faster and extract strategic insights without manual reading.

Approach

Built a hybrid RAG engine (ChromaDB + SQLite). Extracted 2,000+ records from 34 companies × 3 years via SEC EDGAR. Engineered 20+ industry-benchmarked KPIs.

Impact

Production-ready system answering complex financial questions in seconds. Integrated Langfuse for hallucination eval + LiteLLM for failover.

RAG · LLM · Finance GitHub

Zindi Africa Competition

Predicting Financial Health for 9,600 African SMEs

Problem

SME lenders in Southern Africa needed to segment businesses by financial health to improve credit risk decisions.

Approach

Cleaned 9,618 records × 39 features across 4 countries. Engineered 47 features. Benchmarked 4 algorithms with stratified CV.

Impact

Gradient Boosting: F1 = 0.82 weighted, 78.7% accuracy. FastAPI scoring API deployed for real-time batch prediction.

GBM · FastAPI · Africa GitHub

Zindi Africa Competition · National Impact

Togo National Internet Connectivity - Demand Prediction

Problem

The Togolese government needed to predict household internet uptake across the country to guide FTTH infrastructure investment decisions.

Approach

End-to-end ML pipeline on a national survey dataset (~14,400 rows, ~4,000 features). Strict data quality checks, PCA (350 components) for dimensionality reduction, combined with curated categorical indicators.

Impact

XGBoost selected: ROC-AUC = 0.86 on validation. Competition score: 0.7144. National-scale model guiding connectivity infrastructure priorities.

XGBoost · PCA · National Policy GitHub

Recommendation Systems · Matrix Factorization

MicroLend Recommender - Financial Product Recommendation for African SMEs

Problem

50M+ underserved SMEs across Sub-Saharan and North Africa receive no personalized financial product recommendations from MFIs, driving product-client mismatches and high default rates.

Approach

Benchmarked 5 algorithms: User-Based CF, Item-Based CF with cross-selling rules, SVD/NMF matrix factorization, Neural CF (GMF+MLP, PyTorch). 5-fold CV with MLflow tracking. Best model promoted via champion alias in MLflow 3.14 Model Registry.

Impact

Cold-Start Solver for new SMEs with zero interaction history, with confidence scores and automatic CF handoff. FastAPI REST service + Bootstrap 5 interface for loan officers. 14 passing pytest tests.

Python · PyTorch · scikit-surprise · MLflow · FastAPI · scikit-learn GitHub

Statistical Testing · Bayesian Inference

Marketing A/B Intelligence - Statistical Testing Framework

Problem

An e-commerce team needed statistical proof to decide whether to deploy a redesigned landing page to 100% of traffic.

Approach

Frequentist + Bayesian dual analysis on 294,478 visitors (50/50 split, 6 sanity checks). Bootstrap 95% CI, PyMC Beta-Binomial model. Segmentation via logistic regression across time, device, and user type.

Impact

P(B > A) = 9.9%, expected loss 29x higher for Treatment. Returning users harmed (p = 0.037). Recommendation: do not deploy - 3 independent methods agree.

Python · SciPy · PyMC · StatsModels · Streamlit GitHub

Time Series Forecasting · Energy Analytics

AfriPower Forecast - Electricity Demand Forecasting

Problem

Grid operators in Morocco needed short-term demand forecasts (Day+1 to Day+7) to optimize production and prevent overload across distribution zones.

Approach

Probabilistic system on 60,991 rows × 43 features (temporal lags, rolling stats, weather, cyclic encoding). Benchmarked SARIMA, Prophet, XGBoost, LSTM with MLflow tracking.

Impact

LightGBM: MAE = 0.118, RMSE = 0.204, 80% prediction intervals. SHAP identifies temperature + hour as primary drivers. FastAPI + Streamlit executive dashboard deployed.

Python · LightGBM · SHAP · MLflow · FastAPI · Streamlit GitHub

Technical Projects

Engineering depth
across the full stack

Selected projects demonstrating technical breadth - from AI agents to data platforms.

AI Agents · RAG · Agentic Pipeline

DS Interview Coach - Agentic RAG System

Production-grade agentic RAG for BCG X and McKinsey QuantumBlack interview prep. 900+ chunks from Glassdoor, Reddit, YouTube, and real OA screenshots. 4-agent pipeline: Retriever (HyDE + multi-query + RRF + cross-encoder rerank), Cross-Checker (conflict detection), Generator (streaming), Critic (quality gate 7.5/10 with auto re-retrieval). Fully traced in Langfuse. Ragas evaluation pipeline with regression gate.

LangChain FAISS BM25 Groq Langfuse Ragas FastAPI React

GitHub

AI Agents · RAG · Multilingual

AFCON 2025 Morocco - Intelligent AI Assistant

Production-ready multilingual AI assistant for the Africa Cup of Nations 2025. ReAct agent (LangChain) combining RAG on local knowledge base (ChromaDB) + live web search (SerpAPI) + weather API (Open-Meteo) + webpage visitor tool - answering any question about teams, schedules, stadiums, hotels, pharmacies in real time.

LangChain RAG ChromaDB FastAPI React TypeScript Gemini i18n

GitHub

Data Engineering · Analytics

Job Market Intelligence Platform - Skills Really In Demand

Automated scraping of Morocco data job listings (Indeed, ReKrute) with HDBSCAN clustering to segment archetypes and quantify skill demand. Power BI dashboards delivering prioritized skill gap recommendations.

HDBSCAN Power BI Scraping

GitHub

Data Engineering · Azure

Real-Time Food Delivery Analytics Lakehouse

Medallion lakehouse on Azure Databricks with real-time CDC via Debezium + Kafka and SCD Type 2 historical tracking. Optimized Delta tables via partitioning + Z-ORDER for scalable reporting workflows.

Databricks Kafka Delta Lake

GitHub

AI Agents · MCP

DevAssist - AI Agent Platform

Production-ready AI assistant combining LangGraph agents with real-time tool usage via Model Context Protocol. Integrated GitHub, Supabase, Exa MCP servers. Evaluated with DeepEval. Deployed on GKE with Terraform + Prometheus.

LangGraph MCP Kubernetes GCP

GitHub

Data Engineering · Healthcare

Healthcare Data Platform & Analytics Stack

Scalable healthcare data platform with ETL/ELT pipelines (Airflow + Airbyte), multi-layer dbt models (bronze/silver/gold), data quality checks with Great Expectations, and interactive dashboards in Superset.

Airflow dbt Airbyte Superset

GitHub

Computer Vision · Deep Learning

Story Vision - AI Image Captioning

Full-stack AI app generating descriptive captions from images using InceptionV3 encoder + LSTM decoder with attention mechanism. FastAPI backend with async file handling, React frontend with drag-and-drop and dark mode.

PyTorch InceptionV3 LSTM FastAPI React

GitHub

Turning messy data into
decisions that move
businesses forward

Engineer who thinks
like a consultant

Work that speaks
in business outcomes

Why PowerCo's Customers Churn - and What To Do About It

Problem

Approach

Impact

NovaMart Retail Intelligence - End-to-End Advanced Analytics

Problem

Approach

Impact

AI Financial Analyst for Global Finance Corp.

Predicting Financial Health for 9,600 African SMEs

Togo National Internet Connectivity - Demand Prediction

MicroLend Recommender - Financial Product Recommendation for African SMEs

Marketing A/B Intelligence - Statistical Testing Framework

AfriPower Forecast - Electricity Demand Forecasting

Engineering depth
across the full stack

Engineering track with mathematics
and data foundations

Ecole Nationale des Sciences Appliquées d'Al Hoceima

Ecole Nationale des Sciences Appliquées de Fès

Full-stack data capability

Verified credentials

BCG X Data Science Simulation

BCG GenAI Consulting Simulation

Oracle Cloud GenAI Professional

Spark, Hadoop & Snowflake - Big Data Specialization

AWS Cloud Quest: Cloud Practitioner

Neo4j Certified Professional

AI Agents Fundamentals

Let's talk data
& strategy

Turning messy data into decisions that move businesses forward

Engineer who thinkslike a consultant

Work that speaksin business outcomes

Why PowerCo's Customers Churn - and What To Do About It

Problem

Approach

Impact

NovaMart Retail Intelligence - End-to-End Advanced Analytics

Problem

Approach

Impact

AI Financial Analyst for Global Finance Corp.

Predicting Financial Health for 9,600 African SMEs

Togo National Internet Connectivity - Demand Prediction

MicroLend Recommender - Financial Product Recommendation for African SMEs

Marketing A/B Intelligence - Statistical Testing Framework

AfriPower Forecast - Electricity Demand Forecasting

Engineering depthacross the full stack

Engineering track with mathematicsand data foundations

Ecole Nationale des Sciences Appliquées d'Al Hoceima

Ecole Nationale des Sciences Appliquées de Fès

Full-stack data capability

Verified credentials

BCG X Data Science Simulation

BCG GenAI Consulting Simulation

Oracle Cloud GenAI Professional

Spark, Hadoop & Snowflake - Big Data Specialization

AWS Cloud Quest: Cloud Practitioner

Neo4j Certified Professional

AI Agents Fundamentals

Let's talk data& strategy

Turning messy data into
decisions that move
businesses forward

Engineer who thinks
like a consultant

Work that speaks
in business outcomes

Engineering depth
across the full stack

Engineering track with mathematics
and data foundations

Let's talk data
& strategy