Built fromthe inside out.

Final-year internship with the Architecture & Integration Management team of the IoT Cloud Factory. Brief: build a proof of concept assessing whether Hager's API services could be deployed securely on Azure Kubernetes Service over a private network.

• Owned the end-to-end design of an Azure proof of concept assessing secure IoT service deployment on AKS, from reviewing the existing setup through to validation testing.
• Designed a private Azure architecture split across three virtual networks, using VNet Peering and Private Endpoints to secure traffic between AKS, virtual machines and Azure SQL.
• Containerised and orchestrated .NET IoT APIs with Docker, Docker Compose and Kubernetes; added health checks and resolved Azure SQL connectivity issues.
• Worked around an ARM64 incompatibility in the Docker images by setting up a compatible builder, then an AMD64 node suited to the workloads, plus cluster autoscaling.
• Produced a monthly cost estimate for the target architecture — compute, virtual machines, private networking and messaging — stating the assumptions, limits and items still to refine.
• Built an IoT-based energy prediction prototype within a cross-functional team, presented to a Hager-Microsoft jury at the GenIA hackathon.

Verified against a primary source: final-year internship report

Professional internship in an R&D department. Brief: contribute to the system design and build a mobile mapping interface for an industrial inspection system, combining LiDAR, SLAM and real-time visualisation on an embedded platform.

• Designed and validated an embedded real-time 2D mapping prototype end to end, from requirements analysis through SLAM solution selection to functional testing.
• Defined the ROS Noetic / Python software architecture around an RPLIDAR A1M8, an HWT905 orientation sensor and Hector SLAM; structured real-time data exchange over ROS topics with RViz visualisation.
• Built a Python/Tkinter HMI automating LiDAR and SLAM startup, live map display and map saving, so as to simplify handling and reduce operator error.
• Improved localisation reliability by filtering LiDAR data and using motion information; reduced the offsets and drift observed during trials for steadier real-time positioning.
• Called the migration from a Raspberry Pi 4 to an Intel Alder Lake-N100 mini-PC after identifying an incompatibility with the required ROS version, securing software compatibility and smoother execution.

Verified against a primary source: professional internship report

Solo design of a SaaS platform for creating virtual-influencer content. The repository documents an advanced application architecture, but several prerequisites remain before a full SaaS launch can be claimed.

• Led the design of a multimodal SaaS platform single-handedly, covering the whole content creation cycle: images, reusable avatars, scene merging, video, lipsync, media library and administration.
• Defined a serverless architecture separating a multi-page Vite front end from the sensitive services run by Supabase: authentication, PostgreSQL with row level security, storage and fifteen Deno Edge Functions.
• Designed multi-provider orchestration of the AI workloads across fal.ai, Alexya and EvoLink, with asynchronous tracking of video generations, result persistence, gallery and credit management.
• Structured the monetisation model around credit packs and the Stripe Checkout, customer portal and webhook flows, backed by SQL migrations dedicated to billing, transaction traceability and security hardening.
• Industrialised the delivery chain with GitHub Actions, JavaScript/Deno quality checks, Supabase and Cloudflare Pages deployments, Sentry monitoring and analytics consent management.

Verified against a primary source: private repository plateforme-persona — 201 commits, 15 Edge Functions, 33 SQL migrations

A text classification SaaS backed by a detection model hosted on Hugging Face, with authentication, credits and billing.

• Designed a serverless SaaS architecture on React/Vite and Netlify Functions, separating the detection, billing, session management and account deletion responsibilities.
• Integrated an inference pipeline based on RoBERTa via Hugging Face, normalising the score into a probability the interface can use and handling model downtime.
• Structured the usage cycle around Supabase profiles, credits and quotas, checking analysis entitlements up front and persisting the results.
• Implemented the Stripe billing cycle — subscription, customer portal and webhook synchronisation — inside dedicated backend functions.
• Automated front-end build validation through GitHub Actions on every push and pull request to the main branch.

Verified against a primary source: private repository checkai-app — 57 commits, 6 serverless functions

A privacy-oriented PDF editor running every operation directly in the browser, with no application server handling the documents.

• Designed a privacy-by-design architecture running all PDF processing in the browser, with no document transfer to an application server.
• Built a complete document manipulation chain: merge, reorder, rotate, extract, annotate, insert content, sign, watermark and visually redact.
• Assembled PDF.js, PDF-Lib, Fabric.js and Tesseract.js into a single environment to combine rendering, transformation, graphic layers and local OCR.
• Handled internationalisation and responsive design in French, English and Spanish, alongside help flows, tutorials and editorial content.

Verified against a primary source: repositories smartpdf and goeditpdf-public — 40 commits

Preparatory research work on predicting battery state of health and remaining useful life, carried out on a simulated dataset inspired by the NASA PCoE data.

• Designed and compared two Bayesian approaches to state-of-health estimation: a per-battery model and a hierarchical model exploiting information shared across a simulated fleet.
• Formalised degradation dynamics with an exponential law and characterised the posterior distributions of initial state, degradation rate and observation noise.
• Ran NUTS inference with PyMC and NumPyro, posterior predictive checks, and remaining useful life estimation with 90% credible intervals.
• Quantified between-battery variability across eight simulated units and 150 cycles, then analysed the partial pooling effect of the hierarchical model on individual parameters.

Verified against a primary source: public repository bayesian-soh-batteries — simulated data, not validated on real cells

A React/TypeScript prototype built for a Gemini hackathon to assist the analysis and preparation of contract negotiations. The tool offers suggestions and does not replace legal advice.

• Built a contract decision-support prototype accepting text, image or PDF and using Gemini to structure the analysis of risks and sensitive clauses.
• Defined a JSON output schema turning the model response into a risk score, estimated financial exposure, counter-proposals, negotiation difficulty and three-tier strategies.
• Translated the AI results into a decision dashboard combining a risk gauge, clause comparison, negotiation scripts, email options and a printable report.
• Extended the experience with a context-aware conversational mentor and a text-to-speech function for rehearsing negotiation exchanges.

Verified against a primary source: public repository LexiNegotiate- — on-screen statistics are simulated for the demo

Academic project in a team of five: a working connected-door prototype combining face recognition, electric strike control from a Raspberry Pi, and an Android app managing permissions, time windows, history and remote unlocking.

• Co-designed the EyeGuard architecture, a connected-door prototype pairing computer vision, an embedded system, an Android app and MQTT communication, within a 140-hour Agile project.
• Took charge of the face recognition component in a pair: assessed the options compatible with the Raspberry Pi's resources, settled on OpenCV and resolved the installation and configuration issues with the Pi Camera V2.
• Implemented an embedded vision pipeline combining image capture, face detection with Haar Cascades and face embedding generation through a CNN feature extractor.
• Built the detection logging: capture, timestamping and storage of faces, with deduplication over a 50-second window to limit repeated records of the same person.

Verified against a primary source: EyeGuard technical report

A personal object-detection project applied to waste sorting, trained on Google Colab with the public Garbage Detection — 6 Waste Categories dataset.

• Designed and trained an Ultralytics YOLO model through transfer learning to detect six waste categories: biodegradable, cardboard, glass, metal, paper and plastic.
• Set up a reproducible Google Colab pipeline covering dependency installation, runtime Kaggle authentication, dataset download, GPU training and checkpoint saving to Google Drive.
• Evaluated the model after 30 epochs at mAP50 0.496 and mAP50-95 0.344 on the validation set.
• Tested the model on personal photographs and analysed the misclassifications caused by domain shift, lighting, framing and class imbalance.

Verified against a primary source: public repository smart-waste-detector-yolo

Incremental build of an ATS/CRM for independent recruiters and small HR teams, starting from a working prototype and migrating towards a Next.js and Supabase SaaS foundation. The product is explicitly still under construction.

• Led the product design of an ATS/CRM for independent recruiters, structuring the candidate, client, assignment, follow-up and Kanban pipeline workflows around day-to-day recruitment tracking.
• Evolved a working prototype with local persistence into a SaaS foundation on Next.js, React and TypeScript, while preserving the business flows already validated.
• Structured the application around specialised views — candidates, clients, assignments, next actions, search, dashboard and export — to unify activity management.
• Hardened the local Supabase environment under Docker with configuration contracts, controlled key synchronisation, strictly local reset operations, smoke tests and PostgreSQL linting.

Verified against a primary source: private repository CRM_RH — 18 commits, foundation under construction

An MVP helping tradespeople and construction firms prepare quotes from a description of the work or a photograph of the site.

• Built a domain MVP turning a description of the work or a site photograph into a first structured quote, assisted by Gemini.
• Modelled the quote preparation cycle across line items, categories, quantities, pre-tax prices, VAT, discounts, deposits and contracting party details.
• Separated the AI processing from the front end behind a Node.js API protecting the Gemini key and constraining request validation and image size.
• Delivered the output flow through browser printing and PDF saving, then prepared serverless execution on Vercel around a dedicated generation endpoint.

Verified against a primary source: public repository Devis-Chantier-AI — MVP, 4 commits

A full-stack prototype for email open tracking, pairing an API, a web dashboard and a Chrome extension for Gmail.

• Defined a multi-component product architecture pairing an Express/MongoDB API, a React dashboard and a Chrome Manifest V3 extension to embed tracking inside Gmail.
• Designed the core tracking system around unique identifiers and open pixels, recording events and surfacing the indicators in the dashboard.
• Structured the cross-cutting services: JWT authentication, per-plan quotas, real-time notifications over Socket.io and Stripe payment.
• Covered the critical behaviours with Jest/Supertest tests on authentication, tracker creation, free-plan limit enforcement and open counting.

Verified against a primary source: private repository emailtrack — prototype, no public deployment

Part-time fixed-term contract for the 2023/2024 school year, capped at 160 contractual hours, under regional schemes for academic success and widening access to higher education.

• Supported secondary and high school students in Lille through academic tutoring sessions across the 2023-2024 school year.
• Adapted the support to two distinct education levels in order to build students' understanding and independence.
• Contributed to an academic success scheme backed by the Hauts-de-France Regional Council, within a framework planned by JUNIA.

Verified against a primary source: JUNIA tutoring contract

Fifth-year academic project run over eight weeks for Institution Saint-Jacques in Hazebrouck, a real client with roughly 1,200 students from secondary school to post-secondary vocational programmes.

• Led, within a cross-disciplinary team of six engineering students, the scoping of a careers guidance platform, from clarifying the stakes to formalising a solution for future development.
• Structured the user research by cross-referencing the client's expectations, feedback from teachers, parents and students, and a 40-response survey, so as to ground the prioritisation of flows and features.
• Assessed the main careers guidance products on the market and established a bespoke functional positioning, centred on supporting students and centralising content, experiences and events.
• Organised delivery over eight weeks across five phases, a shared workspace and weekly reviews with the client, to maintain alignment and fold feedback into design iterations.
• Translated the prioritised needs into a multi-profile Figma mockup and a functional and technical specification detailing the architecture, navigation flows and expected interactions.

Verified against a primary source: Vis'Avenir report

Academic project in a pair: taking over a non-working drone built by a previous cohort, then studying, redesigning and assembling a quadcopter for visual monitoring of a workshop production line.

• Led the end-to-end redesign of an industrial supervision quadcopter in a pair, starting from a non-working prototype, from clarifying the need through to integration testing.
• Structured the system architecture and specification by translating indoor flight, stability, payload and visual tracking constraints into the design choices of a Quad-X configuration.
• Sized the propulsion and power chain, then selected the 1000 KV brushless motors, 30 A ESCs and 3S 11.1 V LiPo battery in line with the prototype's needs.
• Ran the mechanical design in SolidWorks, the additive manufacturing and the frame assembly; used the first test results to steer the design towards a more compact architecture.
• Defined and integrated the electronics around an F722 flight controller, a Taranis/FrSky link and an ESP32-CAM; configured Betaflight and diagnosed the gyroscope orientation and motor synchronisation issues limiting stability.

Verified against a primary source: drone project report

One-month observation placement on the Al Boustane R1 urban development and hydraulic infrastructure site, delivered for the client Al Omrane. Roadworks, sanitation and utility networks project with an announced value of 78.4 million dirhams.

• Analysed the operational organisation of an urban development site, studying the coordination between works management, site supervision, surveying, the laboratory and the field teams.
• Studied the execution chain for earthworks and roadworks, from topographic setting-out through excavation, cut and fill, trenches, carriageways, parking and pedestrian paths.
• Examined the integration of wastewater and stormwater, drinking water, electricity and telecommunications networks, and the role of survey, setting-out and as-built records in site monitoring.
• Documented the full process for building inspection chambers, from rebar to final concreting, together with the 30-minute watertightness test protocol and its leak detection threshold.

Verified against a primary source: site placement report

A five-day leadership placement run by the French Air and Space Force Academy for JUNIA engineering students, as an immersion in a military environment.

• Took part in an intensive five-day placement in a military environment, run by the Air and Space Force Academy as part of the JUNIA Smart Challenge.
• Tested self-control, resilience and leadership through command workshops and demanding team activities.
• Discovered the values and traditions of the Air and Space Force through field activities — topography, obstacle course and assault course — and unit visits.

Verified against a primary source

Academic smile detection project built from an image dataset sourced on Kaggle, trained on the school's GPU infrastructure.

• Built a computer vision and deep learning project on smile detection from a Kaggle image dataset.
• Trained the model on the HEI/JUNIA GPU infrastructure.
• Worked on a large dataset; the exact volume is recalled from memory and remains to be confirmed before any figure is claimed.

Transitional employment after graduating, alongside the development of personal technical projects.

• Handled in-store customer service under sustained footfall, in a shopping centre.
• Worked as part of a team against commercial targets, to Inditex brand standards.
• Managed difficult customer situations — returns, complaints — diplomatically.

Transitional employment while looking for an engineering role, alongside personal technical projects.

• Covered aisle, checkout and restocking duties in a high-traffic hypermarket.
• Followed strict operational processes — traceability, hygiene, stock rotation.
• Coordinated directly with department managers and the team at a demanding pace.

• Selected among the applicants to write and deliver a talk in front of more than 100 people.
• Worked on narrative structure and on making complex topics accessible.
• Built genuine confidence in public speaking, since reinvested in technical presentations and demos.