import torchdef train():model.fit(){ }loss.backward()pip installtorch.nnoptimizer.step()

Hi, I'm

Salma Ouardi

|

ML Engineer focused on NLP & LLMs. I build retrieval and language systems that go from prototype to production — used by lawyers, doctors, and construction teams.

// 01. about

About Me

Portrait of Salma Ouardi

I'm a Paris-based ML engineer working on NLP and LLM systems. At OnePoint, I build legal and construction AI — RAG pipelines, evaluation, and retrieval optimization for systems used by real end users. Before that, I was part of the NLP team at Ryte AI (healthtech), where I contributed to entity extraction, intent classification, and data pipelines. I'm still early in my career and learning a lot — what I care most about is building things that actually work in production, not just in notebooks.

3

Years in ML

NLP & LLMs

Focus

Healthcare & Legal

Domains

// 02. experience

Experience

ARIA

Built a legal assistant using advanced retrieval with grounded citations over large legal corpora.

Improved reliability and traceability for high-stakes legal analysis.

RAGElasticsearchPromptflowEvaluation

Site Reporter

Designed a voice-to-report assistant that turns field notes into structured site reports for construction teams.

Shipped an end-to-end pipeline from STT to DOCX generation with human-in-the-loop options.

Speech-to-TextFastAPIStreamlitAutomation

DocuScore

Implemented a document scoring engine to prioritize incomplete or high-risk files with model signals and business rules.

Made corpus quality and retrieval-readiness measurable before production rollout.

EmbeddingsSimilarity SearchScoringPython

// 03. skills

Skills

Core

PythonSQLBashLinux

LLM & NLP

RAGLangChainLangGraphPrompt EngineeringEmbeddingsspaCyElasticsearch

ML / Deep Learning

PyTorchTransformersHuggingFaceScikit-learn

Models

GPT-4oMistral 7BLLaMA 3RoBERTa

MLOps & Infra

FastAPIDockerMLflowDVCvLLMCI/CD

Cloud

Azure OpenAIAzure DatabricksGCP Vertex AIBigQuerySpark

// 04. projects

Projects

Production systems I've built at work, and side projects where I explore ideas end-to-end.

Work

ARIA

2025–Present

OnePoint × VINCI Construction

Production Legal RAG system helping lawyers assess contractual risks in tenders and construction contracts.

RAGElasticsearchAzureLangChainPromptflow

Highlights

  • Led retrieval and answer robustness optimization for high-stakes legal documents
  • Built async architecture: Azure Service Bus + Promptflow + Azure Functions
  • Designed evaluation pipelines for legal reasoning precision

Site Reporter

2025

OnePoint × VINCI Construction

End-to-end AI assistant replacing manual construction site reporting. Workers speak, the system generates structured DOCX reports.

Azure OpenAIMistralFastAPISpeech-to-text

Highlights

  • Full voice → structured extraction → DOCX generation pipeline
  • Modular system: Streamlit frontend + REST API + LLM services
  • Validated with real domain users on construction sites

DocuScore

2025

OnePoint

Internal tool for scoring document corpus RAG-readiness — detects duplicates, outliers, and structural anomalies before deployment.

EmbeddingsVector SearchPythonFastAPI

Highlights

  • Embedding + similarity pipelines for large-scale document evaluation
  • Custom anomaly detection: duplicates, outdated docs, outliers
  • Scoring algorithms for structure, quality, and retrieval suitability

Personal

ArXiv Research Copilot

2025

Advanced RAG system for ArXiv paper search with semantic retrieval, citations, and multi-modal support.

RAGLangChainChromaDBOpenAIFastAPI

MediGuide

2024

Fully local RAG assistant answering medical queries from drug PDF documents — no API calls, no data leaves the machine.

LangChainFAISSOllamaDockerCI/CD

FindMyDoc

2024

Intelligent medical query interpreter — classifies intent, extracts entities (symptoms, location, specialties), and links to structured data.

NERTransformersElasticsearchGPT-4HuggingFace

Multi-Agent System

2025

Multi-agent system with tool routing, web search, calculations, and conversational memory — built to learn LangGraph internals.

LangGraphAgentic AIMistralOllama

// 05. education

Education

E1

Université Paris-Saclay

MSc in Artificial Intelligence

2023Paris, France

Specialization: NLP, ML, Deep Learning

Focus Areas

Machine LearningDeep LearningNLP
E2

Ecole des Sciences de l'Information

Engineering Degree in Data & Knowledge

2022Rabat, Morocco

Major: Data Engineering & Knowledge Systems

Focus Areas

Data EngineeringKnowledge SystemsDistributed Data

// 06. writing

Writing

Technical field notes on building AI systems that survive real-world constraints.

Advanced RAG: Lessons from a Production Legal AI System

What I learned optimizing a Legal RAG system for high-stakes document analysis — retrieval strategies, evaluation, and what benchmarks miss.

RAGLLMProduction

2026-04-01 · 1 min read

Read