Finance, data and AI. I work in Artificial Intelligence Solutions at BNP Paribas and develop personal projects in quantitative research and risk analysis.
At work, I combine functional analysis and development for the Group’s procurement teams. In my personal projects, I explore systematic strategies, time series and model validation, with reproducible results.
Two years as an apprentice, followed by a permanent role in the same group.
Artificial Intelligence Solutions
– present
BNP Paribas · Paris · Permanent
Business point of contact for data, reporting and AI in procurement. I combine functional analysis and development, working with the team’s project manager and data scientist.
Analyze business needs and define functional requirements with technical stakeholders.
Develop Python and SQL solutions in Dataiku for the team’s data projects.
Build Python business solutions using LLMs provided by the Group, and contribute to their deployment and optimization.
Develop proofs of concept (POCs) and minimum viable products (MVPs) for AI solutions serving business teams.
Define metrics, build Power BI dashboards and improve reporting tools.
Coordinate business and IT/Data teams to build a shared understanding of the solutions.
ToolsPythonSQLDataikuLLMPower BI
Business Analyst
–
BNP Paribas · Apprenticeship
Worked on the Group’s procurement applications (Ivalua, SAP Ariba) and business intelligence data, combining requirements analysis, technical troubleshooting, data modeling and reporting.
Designed a dashboard to monitor data flows in Dataiku.
Built Power BI dashboards and PostgreSQL queries for business intelligence.
Created UML models and data dictionaries linking technical fields, business definitions and validation rules.
Documented models in Confluence and tracked changes with data and business teams.
Analyzed ServiceNow incidents, reproduced issues in a test environment, and applied SQL fixes and configuration changes within my remit.
Wrote specifications, carried out functional acceptance testing and validated changes with users.
Supported Ivalua and SAP Ariba, managed access rights and worked in English with international teams.
Computer science and business management, alongside an apprenticeship at BNP Paribas. Thesis: the contribution of machine learning to anomaly detection in data flows.
Projects
Featured: quantitative research, credit process analysis, AI business cases and open source.
Filter by skill
Quantitative research · personal project · November 2026
When bonds stop protecting a portfolio: testing a rotation between equities, bonds and gold.
The interactive backtest: performance, drawdowns, correlations and sensitivity to trading costs.
A transparent strategy using four ETFs (SPY, IEF, GLD, SHY), motivated by recent IMF research on weakening equity–bond diversification. It combines momentum, a correlation filter, delayed execution and costs on every buy and sell.
Over 2022–2026, the simulated annualized return reaches 7.6% versus 6.2% for the 60/40 portfolio. However, the maximum drawdown is larger: −24.0% versus −20.9%. The report also examines periods of underperformance and variants that challenge the value of the filter.
An integration with TypeSafe AI’s Jev model classifies macroeconomic notes into structured probabilities. This current analysis is kept separate from historical signals; enabling it on the server requires a connection to the provider.
11 automated tests. Data through November 13, 2026. A retrospective study with no live trading results.
What happens to an option if a share finishes at €120? Change the contract and see who gains, who loses and why.
Exercising the call gives the buyer €20 per share before the premium. After the €8 premium, the gain is €12.
A French-language course for beginners, from a baker fixing the price of wheat to a company exchanging interest payments. Six lessons draw on John Hull, with quizzes and a searchable glossary. An optional lesson inspired by Bouzoubaa and Osseiran compares two price paths that finish at the same level but pay different amounts under a barrier option.
Identify where AI delivers a return, how much, and under what conditions.
The interactive report: 13,087 real applications replayed on a network map.
Process mining applied to a Dutch bank’s real event log: 13,087 consumer loan applications reconstructed step by step. An approved application spends 99.6% of its processing time waiting, and 51% of approved applications require another round of document collection.
Three AI opportunities replayed on real applications, assessed financially through Monte Carlo simulation and evaluated against the AI Act and GDPR. Only one is profitable: checking completeness at submission, with a median three-year NPV of +€48k. The abandonment prediction model does not meet its thresholds and is therefore not deployed.
Includes business analyst deliverables: a scoping brief, AS-IS and TO-BE BPMN models, user stories, an acceptance test plan and change management.
Focus bids on opportunities where you have a better chance of winning.
The opening page: the share of contracts receiving a single bid, by department.
A tender discovery tool for SMEs. A warehouse of 1,264,104 awarded contracts (DECP), built with DuckDB and dbt, powers an explainable hybrid ranking of 7,808 open BOAMP notices, using keywords, embeddings, CPV codes and each buyer’s history.
Since 2024, 25.7% of contracts have received only one bid. SMEs win 59% of contracts but account for 26% of their value. On unseen notices, the hybrid ranking achieves an nDCG@10 of 0.427, ahead of keywords or embeddings alone.
The same business analysis approach: scoping, personas, BPMN, user stories, a business case, compliance and acceptance testing.
Scope, prioritize, assess costs and returns, and manage the risks of an AI use case portfolio.
The showcase: each use case moves through five stations, from scoping to the roadmap.
A web workspace for business analysts and AI committees: a shared scoping form, a value × feasibility matrix with twelve criteria, guided AI Act and GDPR assessment, a business case with Monte Carlo simulation, a three-wave roadmap and an exportable scoping brief.
Across three fictional organizations, 16 use cases assessed, including 4 classified as high risk under the AI Act and 4 with a negative three-year NPV. 195 automated tests.
Find code by what it does, without knowing its name.
Ask a question such as “Where is session expiry handled?” and JevGrep returns the relevant source code snippets, with file paths and line numbers. Use it from the command line, or through an MCP server so a coding agent can explore the repository itself.
A project started with Nassim Arifette, built on TypeSafe AI’s Jev model. My contribution: adding OpenRouter as a third provider for calling Jev, alongside the TypeSafe API and Vercel AI Gateway.
# the provider I added$ jevgrep init --global --provider openrouter
# search for code by its behavior$ jevgrep search --query "Where is session expiry handled?"
Machine learning and anomaly detection in data flows
My master’s thesis examines what machine learning contributes to anomaly detection in data flows. It relates closely to my apprenticeship, where I designed a dashboard to monitor these flows in Dataiku.
Skills
Hackathon · October 2025 · AI Tinkerers
AgentsIRL
A team of five had nine hours to build a multimodal agent that could interpret audio and video streams, using Google Cloud Run and Mistral AI. Although I had never used Cloud Run before, my background in data and machine learning helped me contribute.
Over 115 years, the Gavarnie valley warmed by 2.3 °C and bumblebees shifted their foraging range 118 m uphill, nearly four times less than the rise in isotherms. An interactive story with Python analyses.
A gradient boosting model predicts Indian domestic airfares with an average error of 615 rupees (6.9% of the fare, R² of 0.94). Its 400 trees run in the browser as part of an interactive simulator.
Since 1901, annual rainfall has declined by 29 mm per decade (Sen’s slope, Mann–Kendall test). Yet in August 2018, the state received almost twice its normal rainfall. A three-person project.