A/APPLIED AIENGINEERING PORTFOLIOGet in touch ↗
APPLIED AI · SOFTWARE SYSTEMS · EDGE DEPLOYMENT

Models are only
the beginning.

I build at the intersection of applied AI and software engineering — from experiments and model optimization to systems that make them useful in the real world.

Explore selected work Enter the lab ↓
ESILV × CENTRALESUPÉLEC　·　PARIS, FRANCE　·　PYTHON / TYPESCRIPT
BUILD
MEASURE
DEPLOY
01 / RESEARCH02 / ENGINEER03 / SHIP
FOCUS AREAS

Applied AI · Edge inference · Computer vision · Data systems · Production software

[ 01 — 05 ]
01 / SELECTED WORK

Engineering, in context.

Explore the constraints, decisions and trade-offs behind each project.

01 / 05EDGE AI / DEPLOYMENT
⌁CASE STUDY · 01

From pretrained model to embedded target

Making vision and time-series models useful within STM32 hardware constraints.

PyTorchTensorFlowNVIDIA TAOONNXSTM32
Explore the engineering decisions

Context

At STMicroelectronics, the challenge was to take models beyond the training workstation: adapt them, integrate repeatable workflows and evaluate trade-offs for embedded deployment.

Engineering decisions

  • Integrated Bring Your Own Model workflows into STM32AI-ModelZoo using NVIDIA TAO Toolkit.
  • Worked with PyTorch and TensorFlow transfer learning across detection, segmentation, pose estimation and time series.
  • Explored pruning and quantization with the ST Edge AI toolchain.

Takeaway

Deployment is a systems problem: model quality, memory, latency and compatibility matter together.

02 / 05COMPUTER VISION / SEGMENTATION
◉CASE STUDY · 02

Separating living organisms from debris

Semantic segmentation of large grayscale ZooScan imagery using overlapping patches.

PyTorchU-Net++W&B
Explore the engineering decisions

Context

The work focused on segmenting living organisms from non-living material in large scans.

Engineering decisions

  • Used a U-Net++ architecture with an EfficientNet-B4 encoder.
  • Combined Dice and focal losses for foreground/background imbalance.
  • Reached around 90.4% F1; 92–95% was a target, not an achieved result.

Takeaway

Data pipelines and evaluation matter as much as architecture when labels are scarce.

03 / 05RESEARCH / RECONSTRUCTION
⌖CASE STUDY · 03

Learning nanoscale surface defects

Reconstruction experiments on multiresolution electron microscopy imagery.

PythonPyTorchU-Net
Explore the engineering decisions

Context

At LEM3-LORIA, synthetic data supplemented scarce real microscopy images.

Engineering decisions

  • Generated Perlin-noise height maps with conical and spherical defects.
  • Conditioned reconstruction on lighting and surface normals.
  • The project did not produce a publication.

Takeaway

Synthetic data expands training coverage, but assumptions and failure modes must remain visible.

04 / 05DATA ENGINEERING / TELEMETRY
⌗CASE STUDY · 04

Turning vehicle logs into diagnostics

Aligning GNSS, dead-reckoning and network captures into readable reports.

PythonPandasWiresharkFolium
Explore the engineering decisions

Context

At Stellantis, road-test analysis combined location streams, vehicle references and network traces.

Engineering decisions

  • Parsed PCAP/PCAPNG and NMEA data with Python.
  • Aligned 1 Hz GNSS observations with 10 Hz dead-reckoning data.
  • Built map-based and 3D reports for route behavior and noise metrics.

Takeaway

Useful telemetry tools make time alignment and debugging actions clear.

05 / 05DATA SYSTEMS / STREAMING
↗CASE STUDY · 05

A small system with a full data path

A streaming project exploring event flow, containers and service boundaries.

KafkaDockerKubernetes
Explore the engineering decisions

Context

Tweetoscope explored an event path built around Kafka, Docker and Kubernetes.

Engineering decisions

  • Separated ingestion and processing around event interfaces.
  • Used containers for reproducible environments.
  • Explored orchestration rather than treating the pipeline as one script.

Takeaway

A credible demo exposes system boundaries, not just successful output.

02 / INTERACTIVE ENGINEERING LAB

Touch the trade-offs.

Inspect the demos. Simulated data is labelled honestly; no pretend production metrics.

01 / TERMINAL · LOCAL
portfolio / shell ● LOCAL

portfolio-shell — local terminal

Type help to explore. Commands run locally.

02 / MODEL BUDGET
EDGE AI / RESOURCE EXPLORER

Fit the weights to the target.

Analytical weight-storage estimate; excludes activations, runtime and alignment.

Weight precision
ESTIMATED WEIGHTS2.40 MB
WITHIN BUDGET4 MB
03 / VISION PLAYGROUND
IMAGE PIPELINE / SYNTHETIC INPUT

No upload · no inference request

04 / EVENT STREAM
EVENT PIPELINE / LOCAL SIMULATION

Observe the flow.

EVENTS000
QUEUE00
ERROR RATE0.0%
[09:41:02] ingest event=frame_received
[09:41:02] vision stage=preprocess
[09:41:02] infer model=segmentation_int8

In-memory events only; no Kafka cluster connected.

03 / SYSTEMS THINKING

Follow the whole path.

Model quality is only one part of a deployed system.

SYSTEM MAP / INTERACTIVE

A thin FastAPI service isolates inference dependencies and scales independently.

Constraints first

Memory, latency, energy, connectivity and privacy shape where inference belongs.

Clear boundaries

Separate UI, model artifact, inference service and observability.

Operational by design

Reproducible builds, automated checks and measurable budgets.

Honest evaluation

Distinguish measured results from estimates and simulations.

04 / EXPERIENCE & BACKGROUND

From model to real system.

Engineering training, applied AI experience and a bias toward making the whole pipeline work.

Experience

  1. 01 · Apr — Oct 2025COMPLETED

    Applied AI Engineer Intern

    STMicroelectronics

    Model integration, transfer learning and optimization for STM32 edge AI workflows.

  2. 02 · Apr — Aug 2024COMPLETED

    Connectivity Data Analyst Intern

    Stellantis

    Python telemetry analysis, GNSS/dead-reckoning alignment and road-test reporting.

  3. 03 · Incoming · date to be confirmedINCOMING

    Forward Deployed AI Engineer

    Salesforce

    Upcoming customer-facing engineering role focused on deploying AI-powered solutions in real-world workflows.

Education

Engineering degree

CentraleSupélec

Double-degree engineering programme

Engineering degree

ESILV

Engineering programme · applied mathematics and digital technologies

Exchange semester

Hanyang University · South Korea

Machine learning and data science coursework

WORKING PRINCIPLES

Build for the constraint.
Measure the trade-off.
Ship what proves useful.

HAVE A CHALLENGE IN MIND?

Let's make it work in the real world.

Applied AI, engineering systems or the space between the two.

Explore my GitHub