Soufien HAJJI

Lead Software & AI Engineer · Enterprise AI Copilots in Production · Azure

Thionville, France · Luxembourg area +33 6 69 04 92 63 haji.soufien@gmail.com
Soufien Hajji

Profile

Lead Software & AI Engineer with 9 years designing and operating business-critical C#/.NET systems on Azure, from architecture through production. I take AI assistants from prototype to production operations: deterministic logic in typed code, grounded and cited answers or explicit refusals, automated evaluations that block regressions, and observable deployments. Architecture decisions are documented with rejected alternatives and reversal criteria. ENSTA ParisTech (Information Systems Architecture & Security). Microsoft Certified: Azure AI Apps and Agents Developer Associate. Currently preparing for TOGAF certification.

Professional Experience

Software & AI Tech Lead · CRR3/RWA Programme · Natixis Investment Managers

CRR3 implementation within the risk-weighted assets (RWA) platform: end-to-end ownership of two regulatory-credit workflows; industrialization of the market-risk integration; then design and production delivery of the AI copilot across all three workflows. The platform covers 8 investment portfolios.

Phase 1 · CRR3/RWA programme: solution engineering and market-risk industrialization

Designed and implemented the regulatory-credit capabilities that introduced CRR3 into the RWA application.

  • Designed and built an audited credit-inventory workflow covering ingestion, normalization, historical matching, counterparty identification and data-quality controls, with reproducible checkpoints, rollback and validated enriched outputs.
  • Built the internal and securitized look-through reporting workflow, taking exposures through enrichment, contracts, outstanding amounts and securities to auditable regulatory deliverables.
  • Delivered the application end to end in .NET 8, from workflow and data-processing services to the Razor and SignalR user experience, backed by EF Core and SQL Server.
  • Brought a specialist Python market-risk workflow into the platform behind a FastAPI service, defining the .NET/Python boundary, orchestration and failure isolation while leaving the quantitative calculations under the specialist component's ownership.
  • Took the solution across development, UAT and production on IIS, with AD/LDAP access control, environment-specific configuration, automated build, test and deployment pipelines, and correlated Serilog logs.
.NET 8 / ASP.NET Core MVCRazor / SignalREF Core / SQL ServerPython / FastAPIFull-stack deliveryAutomated testingCI/CD / IISAD / LDAPSerilogRWA · FRTB/SBM

Phase 2 · AI engineering: internal copilot in production across existing workflows

Designed, built and deployed an internal copilot serving three business workflows and 8 investment portfolios, integrated into the existing platform and delivered through its CI/CD pipeline.

  • Built a RAG knowledge system over internal SharePoint documentation and technical and functional specifications: ingestion and extraction with Azure AI Document Intelligence, semantic chunking, embeddings and a vector store, with mandatory citation of the source passage. Users get a sourced answer instead of searching across scattered documents.
  • Enabled on-demand natural-language analysis over production data: translate the business question into an analytical query, execute it through strictly SELECT-only access to the production API, then explain the result. Analysts can interrogate their portfolios without raising a development request.
  • Guided users through the five stages of the credit-inventory workflow and explained the reporting pipeline from enrichment through securities, invoking the existing deterministic services.
  • Applied curated tools, human review, accuracy and faithfulness evaluations, guardrails and production observability. The model plans and explains; deterministic services retain control of workflow state and financial calculations.
LLM agents · tool callsRAG · chunking · embeddings · vector storetext-to-SQL · SELECT-onlyAzure AI Document Intelligenceaccuracy & faithfulness evalsResponsible AI · guardrailsAzure App InsightsPython · SQL · C#

Professional Experience · Previous Roles

.NET Software Engineer, Loans & Securities · Société Générale

  • Developed X-One, a latency- and correctness-sensitive front-office trading platform for the securities lending and borrowing desk, working with trading and operational users in a distributed team of approximately 30 engineers in Paris and 20 in Bangalore.
  • Designed and built the deal position-calculation engine and collateral-valuation features used as key inputs to risk exposure. Delivered prepay-offset processing with runtime-selectable strategies for anticipated interest flows.
  • Led the migration from .NET Framework 4.8 to .NET 6, introduced dependency injection and strengthened SOLID and test-driven practices. Rotated through front-office production support and a quality task force resolving structural defects before delivery freezes.
C# / .NET 6Front-office tradingPosition engineCollateral valuationTDD / SOLIDJenkins

.NET / Azure Consultant · SNCF Réseau

  • Built a real-time operational train-monitoring application connecting stations with maintenance centres, giving each centre a dynamically configured view of its trains, events and operational workflow.
  • Kept the live SignalR view responsive under continuous updates through asynchronous processing, lazy loading and IIS load balancing, across a codebase spanning ASP.NET MVC 5 and ASP.NET Core 3.1 business APIs.
ASP.NET MVC / CoreSignalRReal-time systemsIIS load balancingAsync processing

Azure Consultant / .NET Developer · Orange Business Services

  • Delivered an internal event-ticketing platform as containerized .NET Core microservices behind Azure API Management, with ticket chaining across linked events, secure payment and document-management components integrated with the Box API.
  • Automated build, integration and deployment through Team Foundation Server.
.NET microservicesAPIMDockerTeam Foundation ServerCI/CDBox API / EF

Personal Projects · designed, built & deployed solo

law.soufien.lu · Lex, point-in-time Luxembourg & EU law

Point-in-time Luxembourg and EU law with grounded AI, verifiable provenance and an extensible publisher architecture.

  • Built and operate 2,647 Luxembourg and EU works and collections across 9,374 verified dated versions, behind a publisher-neutral temporal platform.
  • Modeled the metadata catalog on FRBR/JOLux (work, dated consolidation, language expression and manifestation), with controlled publisher, jurisdiction and document-class vocabularies, faceted navigation, explicit gap records when a publisher provides no text, and signed fingerprints binding index, vectors, model and scope.
  • Built and operate a grounded legal research assistant that resolves professional law names and ambiguous intent, plans bounded point-in-time, search, timeline and comparison calls, and asks for clarification rather than silently selecting a law. An evidence ledger and conditional judge verify model-written claims before citations and deep links are rendered.
  • Published a public MCP server exposing ten read-only tools for search, point-in-time retrieval, history, comparison, provenance and coverage.
  • Put local hybrid retrieval behind public relevance and latency holdout gates; the measured benchmark rejected the semantic candidate, so keyword retrieval remains the production default. Recorded the evidence, rejected alternatives and reversal criteria publicly.
C# / .NET 10Legal AI assistantTool orchestrationEvidence judgeMetadata catalogFRBR / JOLuxControlled vocabularyLineage & provenancePublic benchmarks

energy.soufien.lu · Luxembourg Energy Transition Co-Pilot

Grounded solar-PV prospecting combining geospatial analytics, deterministic services and a production AI assistant.

  • Built the public-data and geospatial foundation in PostgreSQL/PostGIS, transforming Luxembourg open datasets into map-based solar-PV prospecting and deterministic calculations.
  • Built and operate a production AI assistant as one bounded agent over five typed tools: NL-to-SQL, document RAG, legal retrieval, tariff calculations and tender discovery. The model plans and explains while deterministic services retain ownership of calculations, read-only data access and rendered actions.
  • Designed the orchestrator as a safety host: it enforces execution limits and typed UI contracts, while a groundedness judge checks model-written prose against the documents and rows actually returned by tools. Direct result cards remain deterministic rather than being rewritten by the model.
  • Implemented hybrid RAG using PostgreSQL, pgvector, full-text search and rank fusion. Used ADRs and evaluations to defer Azure AI Search and reject a premium model that increased cost without improving relevance or groundedness.
Azure OpenAIProduction AI assistantTool orchestrationEvidence judgeRAGpgvector / FTS / RRFNL-to-SQLPostGISEvaluationsTerraform / OIDC

Technical Skills

AI / LLMLLM agents & tool-use (function calling) · Microsoft Agent Framework · RAG (chunking, embeddings, pgvector) · NL-to-SQL · MCP (Model Context Protocol) servers · prompt engineering · evals · Responsible AI / guardrails · Azure OpenAI · Content Understanding · Document Intelligence
LanguagesC# · Python · TypeScript / JavaScript · SQL
BackendASP.NET Core 8 / MVC · EF Core · FastAPI · SignalR · REST & SOAP
DataPostgreSQL / PostGIS / pgvector · SQL Server · Oracle · SQLite / FTS5 · pandas · reproducible pipelines & content-addressed provenance
FrontendReact · Vue 3 / PrimeVue · MapLibre · Vite · Bootstrap
Cloud / DevOpsAzure (Container Apps, Static Web Apps, OpenAI, AI Foundry, Functions, Key Vault, App Insights, APIM) · Terraform (IaC) · GitHub Actions & Azure DevOps · Docker · IIS
PracticesADRs & rejected alternatives · contract & fitness tests · CI-blocking evaluations · AI Act & GDPR governance · mentoring

Certification

Microsoft Certified: Azure AI Apps and Agents Developer Associate

LLM integrationGenerative AI & agentic solutionsPlanning Azure AI solutionsInformation extractionText analysisComputer vision

Earned Jul 2026 · exam AI-103 · verify on Microsoft Learn ↗

In preparation: TOGAF Enterprise Architecture.

Education

ENSTA ParisTech, Engineering degree (Bac+5)

Architecture & Security of Information Systems · 2015-2018

Languages & Interests

French, bilingual · English, fluent (TOEIC 895/990) · German, basics

Coding challenges (9th / 148, Spring Challenge 2020) · Unity game dev · strategy games