Hi, I'm Oussama — I turn 200 GB of noise into a decision.

MSc Big Data & IoT (Mention Très Bien, ENSAM Casablanca). I build data pipelines, ML / DL models, and the dashboards that finally make managers stop emailing me for the number. This page is a notebook — because that's where I think.

Open to roles · Data Analyst · Data Scientist · AI Engineer Casablanca / Rabat / Marrakech · Remote-friendly

In [1]:

cell · imports.py▸ ran in 0.42s

# the usual suspects, plus a few specialists
import pandas as pd
import numpy as np
import torch, tensorflow as tf
from sklearn.ensemble import RandomForestClassifier
from langchain.chains import RetrievalQA      # RAG
import powerbi as pbi                            # DAX, RLS, Time Intelligence
from me import curiosity, rigor, caffeine  # required deps

Out[1]:

	field	value	dtype
0	name	Oussama EL BARAGHI	str
1	target_roles	Data Analyst · Data Scientist · AI Engineer	str[]
2	degree	MSc Big Data & IoT — Mention Très Bien	str
3	school	ENSAM Casablanca · ENS Marrakech	str
4	current_role	Data Analyst de Chantier @ Sogea Maroc	str
5	biggest_dataset	200 GB hyperspectral (UM6P ALAB)	str
6	best_model_acc	0.95 — 0.98	float64
7	languages	['ar','fr','en','amazigh']	list
8	open_to_work	True	bool

1. The 30-second pitch

I sit at the seam between BI / reporting and ML / Generative AI. That means I can clean a messy Excel export and ship a Flask API serving an LLM — and I've done both in the last year.

Coming from a research lab (200 GB hyperspectral) into a construction site (effective KPIs, real budgets), I've learned the same lesson twice: the model doesn't matter if the data isn't trusted. So I start with data quality, then everything else follows.

"Replace the manual report. Trust the dashboard. Then go model something interesting." — what I tell every manager I work with.

In [2]:

cell · summary.py

me.describe(include="all")

Out[2]:

Largest dataset

200 GB

hyperspectral · UM6P

CNN accuracy

95–98%

plant-health classifier

ETL time cut

−30%

Univers Acier reporting

Languages spoken

4

AR · FR · EN · ⵣ

2. Featured projects

Three projects that span the data stack — research-grade DL, production-grade GenAI, and bread-and-butter BI. Numbers are real, methods are documented, leakage was checked.

In [3]:

cell · project_01_chickpea_health.py▸ ran in 4h 12m (full training run)

# Master's thesis @ UM6P Analytics Lab — feb→sep 2025
model = CNN(in_bands=204, n_classes=4)
loader = HyperspectralLoader(root="chickpea/", size="200GB")
trainer.fit(model, loader, epochs=40)
print(f"val_acc = {trainer.best:.3f}  ✓ stakeholder-grade")

Out[3]:

val_acc = 0.973 ✓ stakeholder-grade · F1_macro = 0.961

PROJECT · 01 / 03 · UM6P ANALYTICS LAB · MSc THESIS

Classifying chickpea plant health from 200 GB of hyperspectral imagery

Built the full pipeline — ingestion, validation, CNN in PyTorch, reporting — for a research project that turned drone-captured spectra into a usable plant-disease signal. Wrote the data-quality checks first; they caught corrupt scans before they hit training.

Val. accuracy

97.3%range 95–98

Dataset

200 GB204 bands

Pipeline

End-to-end

Confusion matrix — 4-class plant health

Test set · n = 12,420 patches · diagonal = correct

PythonPyTorchCNNComputer VisionHyperspectralscikit-learnPandasData Quality

In [4]:

cell · project_02_hr_rag.py

# RAG system for resume ↔ job-offer matching
retriever = VectorStore.from_resumes(cv_corpus).as_retriever(k=5)
qa = RetrievalQA.from_chain_type(llm, retriever=retriever)
app = Flask("hr-assistant")
app.route("/match", methods=["POST"])(qa.handle)

Out[4]:

PROJECT · 02 / 03 · GENERATIVE AI · FULL-STACK

Automated HR recruitment assistant — RAG over a CV corpus

A LangChain RAG pipeline that ingests CVs, embeds them, and ranks candidates against a job offer with a justification per match. Served via a Flask REST API so the HR team queries it from their existing tool — no new UI to learn.

Architecture

RAG

Serving

Flask API

Deployable

Docker

System architecture — request → ranked match in 1 hop

Generated from project README · 4 services · async

LangChainLLMsRAGTransformersFlaskREST APIVector DBDockerPrompt Engineering

In [5]:

cell · project_03_univers_acier_bi.py

# Internship @ Univers Acier — automated steel-sales reporting
pipeline = (PowerQuery(sources=["erp", "crm", "csv"])
              .query_fold().deduplicate().to_powerbi())
dax("YoY_Margin", "DIVIDE([margin]-[margin_ly],[margin_ly])")
log("manual reports retired ✓")

Out[5]:

PROJECT · 03 / 03 · UNIVERS ACIER · BI INTERNSHIP

Replacing manual Excel reports with a self-serve Power BI suite

Automated the company's sales reporting end-to-end: SQL extraction, Power Query ETL, dynamic dashboards with DAX (YoY, RFM, margin per SKU). Reduced processing time by 30% and eliminated the weekly reconciliation errors that used to land in the COO's inbox.

Processing time

−30%vs manual

Recon. errors

0eliminated

KPIs tracked

24

Weekly close — minutes from raw export to signed-off dashboard

2024 internship · 13 weeks · before/after automation

Power BIDAXPower QuerySQLETLExcel TCDData Modeling

+ four more, briefly

Out[6]:

	project	stack	headline result
04	Pneumonia detection from chest X-rays	TensorFlow, Keras, CNN	92% accuracy · medical-image classification
05	Predictive maintenance — industrial machinery	Scikit-learn, Random Forest, Pandas	85% failure-forecast accuracy from raw sensor data
06	Expense & Earning Tracker with AI insights	Spring Boot, React, Python, PostgreSQL	Full-stack app · personalized savings recommendations
07	E-commerce performance dashboard	Power BI, DAX, SQL, Pandas	YoY · RFM · stock-profitability KPIs in one report
08	Customer segmentation (retail)	Scikit-learn (clustering), Python	Behavioural segments → targeted marketing & CLV gain

3. Skills, by self-reported confidence

The bar reflects: "I'd be comfortable defending a decision I made with this on day 1." Tools I use daily get a ★.

In [7]:

cell · skills.py

skills.groupby("family").apply(plot_bar)

Out[7]:

Data & analytics

SQL ★

92

Python (Pandas) ★

90

Power BI / DAX ★

88

Power Query

86

Excel (TCD)

84

Tableau

72

R

58

PostgreSQL / MySQL

82

ML / DL / GenAI

Scikit-learn ★

86

PyTorch

82

TensorFlow / Keras

76

LangChain · RAG

80

Prompt Engineering

78

OpenCV · NLTK

70

Engineering & cloud

Flask · REST API

82

Docker

74

Git / GitHub · CI

84

Apache Spark

60

AWS (S3 · EC2 · RDS)

68

Azure

58

4. Timeline

Engineering school → research lab → construction site. The plot twist is that the math is the same.

In [8]:

cell · timeline.py

df.query("actor == 'oussama'").sort_values("date", ascending=False)

Out[8]:

JAN 2026 — PRESENT · BEN GUERIR

Data Analyst de Chantier — Sogea Maroc

Construction · operational analytics

Daily analysis of headcount/attendance/yields per worksite. Built Power BI + Excel dashboards (rendement, coût/m², avancement) that replaced manual reports — managers now decide from the dashboard, not from a forwarded email.

FEB – SEP 2025 · RABAT

Data Science Researcher — UM6P Analytics Lab (ALAB)

Research · 200 GB hyperspectral dataset

Designed the validation & quality-control protocols (first line of governance), built automated Python anomaly checks at ingestion, then trained CNNs in PyTorch reaching 95–98% classification accuracy. Reported results to non-technical stakeholders.

JUN – AUG 2024 · CASABLANCA

Data Analyst / BI Intern — Univers Acier

Steel distribution · BI & reporting

Automated ETL flows in SQL + Power Query → −30% processing time, zero reconciliation errors. Built dynamic Power BI dashboards (DAX) for sales & margin tracking. Wrote complex SQL for ad-hoc analysis and data-quality validation.

2023 — 2025 · CASABLANCA

MSc Big Data & IoT — ENSAM Casablanca

Mention Très Bien · Thesis: Chickpea Plant Health (ML & DL)

2020 — 2023 · MARRAKECH

BSc Computer Science — ENS Marrakech

Foundations · algorithms · databases

5. Certifications & languages

In [9]:

cell · credentials.py

me.verified(include=["certs", "langs"])

Out[9]:

✓

DeepLearning.AI — Supervised Machine Learning

Andrew Ng · Coursera

✓

AWS Academy — Machine Learning Foundations

AWS

✓

Coursera — SQL for Data Science

UC Davis

✓

AWS Academy — Cloud Foundations

AWS

✓

Scrum Foundation Professional Certification

CertiProf

Languages

Arabicnative

Amazighnative

Frenchfluent (C1)

Englishprofessional (C1)

6. The "weaknesses" cell

Recruiters always ask. I let the kernel answer.

In [10]:

cell · weaknesses.py

print(me.weaknesses())

Out[10]:

PerfectionError: I sometimes over-engineer the EDA before checking with stakeholders. File "growth.py", line 14, in improve self.timebox_eda(max_minutes=90) # actively working on this ProductionGap: my biggest wins so far live in notebooks & Power BI — fewer in prod APIs. → shipped the RAG assistant as a Flask service this year, taking an MLOps course next.

7. Hire me — or just say hello

The fastest way is the dock at the bottom: type a question and the notebook answers from this page using a real LLM. Or use the boring channels below.

Oussama EL BARAGHI
elbaraghi.oussama@gmail.com
+212 6 18 13 10 94
Casablanca / Rabat / Marrakech · UTC+1
# end of notebook · kernel idle

↗ Email me ↗ LinkedIn ↗ GitHub ↗ Resume.pdf