Machine Learning Engineer · AI Engineer · Data Scientist

Diego Agudelo

I build machine-learning systems from the maths up to a running service, and I report how each one was evaluated and where it falls short.

Among them: a neural network written from scratch without an ML framework, a salary-prediction pipeline deployed as a public web app, and a question-answering system over French labour law that refuses off-topic questions.

At École 42 Paris I direct the AI Lab of 42AI: I scope its projects, recruit and mentor the students who build them. Before that, three years as a civil engineer with Egis and Sixense, inspecting tunnels, metro lines and buildings.

École 42 Paris · 42AI AI Lab · Egis · Sixense

Now · Updated 26 Sep 2026

Building RAG-Syntec, a retrieval-augmented QA system over the Syntec collective agreement.
Directing the AI Lab at 42AI.
Open to work, available now.

Selected work

Six projects, each with my part, how it was evaluated and its limits.

1.1Weight Your Wage

Predicts developer salaries from the Stack Overflow survey. A 42AI AI Lab team project covering the full path from raw data to a public web app; the model is a deliberate baseline, the pipeline is the work.

Deployed · public web app

raw survey

MinIO

ETL

PostgreSQL

clean labels

Cleanlab

train · Huber

Lightning · MLflow

serve

FastAPI jobs

web app

Next.js

monitoring — Grafana · Prometheus · Loki

Fig. 1 — Pipeline as implemented in the repo. Stages outlined in accent are where most of my commits went.

MY PART: My commits went mostly to the model training code (PyTorch Lightning, MLflow), the FastAPI service, the Docker Compose setup and the monitoring dashboards. I built the Next.js web front end alone, in a separate repository.
RESULT: Deployed and public. No accuracy claim: the model is a placeholder by design.

Read the case study →Code ↗Live demo ↗

CONTEXT: 42AI AI Lab
ROLE: Team of 5
STATUS: Done
STACK: Python
PyTorch
PyTorch Lightning
pandas
scikit-learn
Cleanlab
MLflow
FastAPI
PostgreSQL
MinIO
Docker
Docker Compose
Prometheus
Grafana
Loki

1.2 Multilayer perceptron

Classifies breast tumours as benign or malignant with a neural network written in NumPy: layers, losses and optimisers behind a Keras-like API.

97.4 % val. accuracy
569 samples

96.5–97.4 % validation accuracy across three 80/20 runs (seed 42); the main run with early stopping on validation loss reached 96.5 %.

Case study →Code ↗Notebook ↗

Adam, 24-24, run A97.4%
Early stopping, main run96.5%
Adam, 24-24, run B94.7%

Fig. 2 — Validation accuracy

CONTEXT: École 42 · AI projects
ROLE: Solo
STATUS: Done
STACK: Python
NumPy
pandas
matplotlib

1.3 RAG-Syntec — IN PROGRESS

Answers questions about the Syntec collective agreement and refuses off-topic ones before calling the LLM. 172 documents, 8,447 chunks, dense retrieval on Chroma.

172 docs · 8,447 chunks
55 labelled questions

Threshold calibration only: in-topic questions stayed below a distance of 0.701, off-topic ones above 0.780; the threshold sits at 0.74.

Case study →Code ↗

In-topic, largest distance0.701
Refusal threshold0.740
Off-topic, smallest distance0.780

Fig. 3 — Nearest-chunk distance

CONTEXT: Personal project, started from a technical-test skeleton
ROLE: Solo
STATUS: In progress
STACK: Python
LangChain
Chroma
FastAPI
Ollama
OpenAI API
Docker
Docker Compose
pytest

1.4 CapTech IDF — IN PROGRESS

Maps 79,099 Île-de-France companies from public Insee data across 16 technology domains. A company joins a domain only with a cited, dated source; 23 have one so far, in AI, cybersecurity and embedded systems. The code is private.

79,099 companies
1,343 indexed segments

Sirene stocks

Parquet · DuckDB

load

Supabase

crawl sites

robots.txt · 1 req/s/host

evidence corpus

URL · page · block filters

cited domain

source · date

map · 2D/3D scene

MapLibre · three.js

tests — 1,000+ automated tests (Vitest)

Fig. 4 — Data flow as documented in the private repo. Outlined in accent: the evidence path; a company joins a domain only with a cited, dated source.

Corpus cleaning cut indexed segments from 2,105 to 1,343; 45 of 46 organisations kept business text. Retrieval is benchmarked on 18 queries as positive–unlabelled, so no accuracy or F1 is claimed.

Case study →Live demo ↗

CONTEXT: Personal project
ROLE: Solo
STATUS: In progress
STACK: TypeScript
Next.js
Supabase
MapLibre
ECharts
three.js
DuckDB
Vitest
Cloudflare Workers
Claude Code

1.5 Leaffliction

Classifies apple and grape leaf photos into 8 classes (6 diseases, 2 healthy) with a PyTorch CNN; augmentation balances the classes before training.

94.1 % held-out accuracy
8 classes

94.1 % accuracy on 152 held-out test images.

Case study →Code ↗Live demo ↗

Validation (model selection)98.1%
Held-out test, 152 images94.1%

Fig. 5 — Accuracy

CONTEXT: École 42 · AI projects
ROLE: Team of 2
STATUS: Done
STACK: Python
PyTorch
OpenCV
Albumentations
PlantCV
matplotlib

1.6 DSLR

Sorts Hogwarts students into one of four houses from their course scores, with one-vs-all logistic regression and gradient descent written from scratch — no ML library.

4 houses
13 course features

On the 1,600 training students the model reproduces the house labels with 97.9 % accuracy (34 errors). This is a training-set score, measured on the same data it was trained on — there is no held-out test set.

Case study →Code ↗Notebook ↗

CONTEXT: École 42 · AI projects
ROLE: Solo
STATUS: Done
STACK: Python
Matplotlib

All projects

32 projects, most from the École 42 curriculum. Each domain opens the index filtered on it, with code and demo links.

Open the full index →

Journey

From civil engineering to machine learning.

Director, AI Lab — 42AI, École 42’s AI associationI scope the Lab’s projects, recruit and mentor participants. First shipped project: Weight Your Wage.
2026 –
École 42 Paris — Expert in IT architecture (RNCP level 7)Data track, ends 2026. C and C++ systems programming, then machine-learning projects: DSLR, multilayer perceptron, Leaffliction.
2023 – 2026
Civil engineer — Sixense Engineering, EgisStructural surveys in La Défense, inspections of the A86 Duplex tunnel, feasibility work for Cairo metro line 6, waterproofing control on the RER E extension.
2020 – 2023
Master’s in civil engineering — Université de LimogesTrack: inspection, maintenance and repair of structures (IMRO).
2018 – 2020
Civil engineering degree — Universidad Libre, Pereira (Colombia)Final project: structural recalculation of La Isla bridge, Belén de Umbría.
2011 – 2016