A hiker seen from behind on a forest trail, facing a mountain.

ROMAINSCHRIJVERS

ENGINEERING STUDENT AT INSA LYON · EX-L'ORÉAL INTERN · FREELANCE AI & AUTOMATION

OVERVIEW

Learn the trade first,
then build the tool
it never had.

Computer science engineering student at INSA Lyon, freelance, and a former intern at L'Oréal. I start from a real problem, at a real client, and build the tool that solves it.

This page brings together the track record, the projects, the working method and the recommendations.

Who I've worked with.

JOURNEY

What got me here.

From elite football to a global group's information systems. Four experiences that make up my path.

012020–2025

Five competitive seasons, from Le Mans FC to INSA Lyon.

Four Le Mans FC players lined up on the pitch before a match.

Discipline
High standards
Leadership
Team spirit

Football at a high level alongside my studies, from high school to engineering school.

Le Mans FC academy, captain of the national U17 squad
Coupe Gambardella quarter-final with AS Saint-Priest
Twice French Grandes Écoles champion and winner of the Saint-Gobain Football Cup at Clairefontaine with INSA

022022–2028

INSA Lyon, computer science department.

A student working on a laptop at a table.

Scientific mindset
Problem solving
Fast learning
Time management

A degree that combines scientific fundamentals with applied software engineering.

Web development, data, software architecture and systems
Team and industrial projects, run with collaborative engineering methods
Adapted timetable in the elite-athlete programme during the preparatory years, 2022 to 2025

032024–2026

ETIC INSA Technologies, voted France's best engineering-school student consultancy for 2024–2025.

A team in evening dress on a stage, one of them speaking into a microphone.

Business sense
Client relations
Management
Project steering

A year as account and project manager, then running the sales team: strategy, management and oversight.

The whole cycle of an engagement: prospecting, scoping, pricing and project follow-up
€100k+ of projects signed personally from prospecting
+275% acquisition on the previous year
A team of around 40, in a consultancy with roughly €500k of annual revenue

04June–August 2026

L'Oréal Global Beauty Tech in Paris, Practice & Governance team.

A wall of lit laboratory flasks filled with spices, flowers and wood.

Architecture & integration
RAG & applied AI
API governance
Complex IT landscapes

An IT organisation recentralising after years of decentralisation, an integration platform moving to Workato, and AI raising new questions of adoption, monitoring and control.

Audited and improved the API Developer Portal's RAG chatbot: relevance, context, cost
Mapped the system's interfaces and data flows, with a roadmap for evolving its data model
Data quality and Business API governance, supervising a team in India

SELECTED WORK

What I'm building now.

At ETIC INSA Technologies I signed €100k+ of client work and ran 10+ projects. Now I build them myself, freelance, with 2+ clients delivered so far.

The ARICI application in service: the tender folder screen, on the firm's own workstation.

AI
RAG
DOCUMENT INTELLIGENCE
FULL-STACK

ARICI · Automated public tender responses

Every French public tender opens with a pack of several hundred pages, and it has to be read through before anyone can even tell whether the job is worth bidding for. I designed and built the application that absorbs that reading, then drafts the technical proposal from the firm's own library of past work rather than from generic text.

Analysis: Documents classified, requirements extracted and deduplicated, contradictions and risky clauses raised explicitly.
Domain: Only the structural works package is kept, without ever dropping the interfaces with the other packages: openings, shafts, ducts.
Library: Site references, CVs, equipment sheets and resources vectorised once, reused on every tender. Hybrid retrieval: dense search and French full-text search, merged by Reciprocal Rank Fusion.
Drafting: The structure is derived from the tender's real scoring grid, then each section is written from the passages of the pack and the library that actually matter, with every claim tied back to its source.
Role: Design, architecture and full-stack development, deployed inside the client's infrastructure.

Price, lead time and headcount are still entered by the firm. The tool speeds up the reading and the writing; it does not commit a small business on its behalf.

80 %less time spent on first-pass analysis

FastAPI
PostgreSQL + pgvector
Hybrid RAG
Mistral
React 19

AUTOMATION
CRM
LLM

Propale Bot

At ETIC every commercial proposal was written by hand, then copied again into the contract and into the CRM. I automated the whole chain, working from the archive of the firm's strongest past proposals. That archive is what gives the generated text the level the team expects.

Material: Drafting draws on the archive of the best past proposals, the contractual limits and the in-house methodology.
Generation: Scope, methodology, budget and payment schedule produced from the phasing document.
Review: The proposal stays editable after the client's feedback, before it turns into a contract automatically.
Follow-up: Budgets, schedules and invoicing dates written straight into HubSpot, with no re-keying.

~~4–6 h~~10–30 minfor the whole chain

Python
REST API
HubSpot API
LLM

Screenshot of Propale Bot showing the stepper and the deal validation step.

DATA
POWER BI
EXEC REPORTING

SOCOTEC · Market steering tool

This was not a technical problem first. It meant understanding how SOCOTEC reads its own commercial performance against the construction-inspection market, and which indicators genuinely count. I designed the data model, then a report where the group's executive committee sees the whole picture and drills into the detail wherever it needs to.

Domain: Permitted work and started work set side by side: the reading that separates potential from committed.
Model: The Sitadel 2 API for the market, internal data lake and ERP for activity, reconciled into a star schema with DAX measures.
Navigation: The group's executive committee gets the overview and can drill anywhere: territory, business segment, period. The regional and branch views extend the same reading.
Outcome: A single report running from national level down to the branch, ready for deployment in the group's Power BI environment.

The work was not showing everything, it was choosing what stays visible at each level.

Power BI
Power Query
DAX
Sitadel 2

See every project

METHOD

Two loops. One direction: the field.

The short loop I close on my own: enough of the client's material gathered up front to judge my own output. The long one only the client can close, and what comes back corrects my understanding of their trade, not a queue of tickets.

01Scoping
One conversation, the questions nobody asked, then a scoped and priced proposal within 48 to 72 hours: what will be done, what will not, and on what terms. So far, 100 % of the engagements taken on have been delivered and signed off.
02Context
I don't wait until I've mastered the trade to start. I gather its material, the client's own documents, the rules, the edge cases, targeted research, until I can make progress without asking them about every decision. At ARICI, that is what surfaced that the binding deadline sits in the contract award form, not in the general conditions.
03Build
A weak first version asks the client to take it on faith; a solid one lets them judge. So it has to cover most of what was expected already, with enough substance to be used, assessed and argued with.
04Evaluation
I grade my own output, check what holds and what does not, and go again, without taking up the client's time. At ARICI, an LLM judge replaced a threshold that kept getting it wrong, and whatever stays uncertain is flagged rather than invented.
05Field
A version people can argue with goes further than a specification people approve. Handling it in their own environment is what makes the gaps visible, and those gaps go back into the material for the next round.

SELF-EVAL

The short loop

Build, grade, correct, go again. It turns fast because it depends on nobody else: the product closes in on what was expected before the client is ever asked.

CLIENT FEEDBACK

The long loop

It runs through the field, comes back to the context, then feeds into the product again. What the client corrects joins the reference material: every round sharpens the understanding of the problem and makes the next one more accurate.

RECOMMENDATIONS