Hello, I'm

Haroun Ben Salem

Data engineer & AI engineer

I build the quiet systems data lives in — Airflow orchestrations, Kafka streams, dbt-tested models — and the AI products that read them.

Seeking a high-impact PFE — 2027

04engineering internships
07systems shipped
03AI products in the open

Systems that hold · Models that answer · PFE 2027

01 · Proof of work

Résumés say data engineer.
This page shows the pipelines.

Every number on this page belongs to a system I actually built — with names, dates and gradeable output. Read it the way a manager reads a README: context, challenge, build, proof.

— the working method behind every section below
04
Engineering internshipsBI → cloud automation → applied AI
07
Systems shippedeach one a working pipeline behind it
03
AI products in the openrecommenders, RAG, ops agents
01
Open ticketa high-impact PFE 2027

02 · Selected work

The receipts.

One flagship case, four working systems. Every screenshot below is the real product surface — the generated stand-ins retired.

The InvoiceGuard operations cockpit on the synthetic billing corpus: review queue, quality gates and the current run's scored invoices

Flagship — full case · InvoiceGuard

Intelligent utility-billing audit & anomaly detection system.

Role

Design & build, solo

Stack

Python · SQL · Pandas · FastAPI · PostgreSQL · XGBoost · Great Expectations

Status

Engineered end-to-end

01

The context

Utilities bill through SAP IS-U and Opencell. Two export schemas, bending tariffs, and errors that compound into retries, VAT drift and bills customers should never receive.

02

The brief

Standardize the heterogeneous billing exports into one canonical model, encode the rules a real audit would apply, and let statistics catch what rules can't.

03

The build

23 deterministic billing rules across tariff compliance, consumption anomalies and VAT reconciliation — plus 38 Great Expectations-compatible data-quality checks. An XGBoost anomaly engine with hybrid risk scoring, batch and online at parity. FastAPI services on PostgreSQL with SQLAlchemy and Alembic migrations. Factur-X / EN 16931 e-invoice validation. Prometheus metrics, GitHub Actions CI.

  • 23billing rules encoded
  • 38data-quality checks
  • Paritybatch / online scoring
  • EN 16931e-invoice validation
The VIF Cosmetics homepage: your skin story told through a real AI skin-analysis product
01

VIF Cosmetics

Shipped · in production — Celestial Wave Digital

AI skincare analysis that reads a face and recommends like a pharmacist.

Facial images drive a skin-profile analysis in a multilingual embedding space; retrieval fuses pgvector semantic search with PostgreSQL full-text over the catalog; a deterministic recommender ranks products by goals, eligibility and weighted rules — with the analytics that grade every suggestion.

Next.js · FastAPI · Python · PostgreSQL · pgvector

The RAG e-commerce assistant: LangChain-orchestrated retrieval over a 10,000-product catalog with FAISS vector search
02

RAG e-commerce assistant

Personal system — evaluated

Answers with receipts over a 10,000-product catalog.

LangChain-orchestrated retrieval with FAISS vectors, served through FastAPI on Docker — and graded against a held-out evaluation set instead of vibes.

  • 92%answer accuracy
  • <1sp95 latency

LangChain · OpenAI · FAISS · FastAPI · Docker

Stylized render of the fraud-detection event stream: five thousand events per second flagged and scored

Stylized render of the event stream.

03

Real-time fraud detection

Streaming — live

Five thousand events a second, judged once.

Kafka ingestion into Spark Streaming, an Isolation Forest anomaly model, PostgreSQL persistence and live Grafana dashboards watching the drift.

  • 5K/sevents ingested
  • F1 0.94Isolation Forest

Kafka · Spark Streaming · PostgreSQL · Grafana

Banking data warehouse: raw data through dbt dimensional modeling to executive dashboards
04

Banking data warehouse

Dimensional modeling — dbt-tested

Numbers a CFO can sign.

Dimensional model on PostgreSQL with dbt Core transformations and tests, SCD Type 2 customer tracking, automated ELT on GitHub Actions, and Metabase dashboards for deposits, loan performance and customer behavior.

  • SCD 2customer history
  • dbttested transforms

dbt Core · PostgreSQL · Python · Metabase · GH Actions

✦ The rest of the shelf

  • Customer segmentation — K-Means, GMM & Deep Embedded Clustering over RFM features; PCA / autoencoder reductions; shipped as a Streamlit app.
  • Marketing data pipeline — daily Airflow extract-clean-load orchestration with Grafana-monitored data-quality SLAs and Slack alerts.
  • Gesture recognition — a custom OpenCV CNN: eight gesture classes at 96% accuracy, live at 30 FPS.

03 · Track record

Four internships,
four different rooms.

  1. 2026 → now

    The VIF Cosmetics product chrome: the VIF wordmark, the Analyse navigation and the pink campaign florals — a different frame from the homepage

    Celestial Wave Digital — AI Engineer / Data & AI Engineering Intern

    Shipping VIF Cosmetics end-to-end: the AI skin-analysis workflow, hybrid pgvector retrieval, a deterministic recommendation pipeline, the REST workflows behind it — and the analytics that grade every recommendation.

  2. 2025

    AI agents automating OpenStack diagnostics: log analysis, VM monitoring and health checks

    FelCloud — AI Engineer Intern

    LangChain + Ollama (Mistral) agents automating OpenStack diagnostics; log-analysis workflows that cut system response time; modular agents for VM monitoring and health checks.

  3. 2023

    Power BI dashboards steered by ETL pipelines: total sales up 12.5 percent across regions

    Khallasli Sousse — BI Analyst Intern

    End-to-end ETL pipelines and automated refresh workflows that minimized manual reporting hours; dynamic Power BI dashboards that department heads steered decisions by.

  4. 2022

    Node.js real-time chat application: Socket.IO messaging across a MongoDB and MySQL hybrid store

    PayPOS Sousse — Software Engineering Intern

    A real-time chat application on Node.js + Express + Socket.IO; a hybrid MongoDB + MySQL database architecture; REST APIs supporting scalable real-time messaging.

Education

  • Engineering Degree in Data Science — EPI Digital School, 2023–2026, in progress
  • Bachelor's Degree in Business Intelligence — Higher Institute of Management, Sousse, 2020–2023
  • Baccalauréat in Mathematics — 2020

Certifications

  • AWS Cloud Practitioner Essentials — Amazon Web Services
  • Professional Certificate — Google / Coursera

Languages

  • Arabic — native
  • English — C2, advanced professional
  • French — C1, professional

The toolkit

Pipelines & data

  • Airflow
  • dbt
  • Kafka
  • Spark
  • PostgreSQL
  • MongoDB
  • Snowflake
  • ETL / ELT

AI & ML

  • TensorFlow
  • PyTorch
  • Keras
  • scikit-learn
  • LangChain
  • Ollama / Mistral
  • Hugging Face
  • OpenCV

Cloud & ops

  • AWS (S3 · EC2 · Lambda)
  • OpenStack
  • Docker
  • Kubernetes
  • Git / CI-CD
  • Linux

Visualization

  • Power BI
  • Tableau
  • Streamlit
  • Grafana
  • Plotly / Matplotlib
Portrait of Haroun Ben Salem

04 · Open ticket

Bring the problem.
I'll bring the pipeline.

Documented work, tested transforms, honest numbers — and the receipts above to check me against. One email starts the conversation.

Haroun Ben Salem's suit in detail — navy lapels, oxblood tie and pocket square, white boutonnière and watch, before painted tiles and a gilded frame

Haroun Ben Salem

Data engineer · AI engineer · Tunisia

Ships end-to-end · Tests the transforms · Prints honest numbers

Base

Tunisia — available for PFE 2027