Finding knowledge in a sea of records.

I’m Loujain Liekah, an AI and data scientist. I carry research into production on real company data: entity resolution, data quality and privacy-safe AI, and the technical choices that make them hold at scale.

The words drift like a tide. Move your cursor to bring them into focus, and click to send a ripple.The words drift like a tide. Tap to send a ripple.

Results you can check

91% cheaperAn open-weight LLM chosen on real data: 16.5× smaller than the first pick, hosted in France.Case study
1.25 billion to 40,000Candidate pairs reduced to about 40,000 comparisons for entity resolution, run in seconds.Case study
Ranked 10thD-HAT, my research method, on the Amazon-Google leaderboard, above GPT-4o and Llama-3.1-70B, without fine-tuning.Case study

Same thing, different records.

Matching records that describe the same product, company or person is decided before the matching itself: by the quality of the data, and by the number of comparisons, which grows with the square of the data. Walk through the steps on real benchmark records and on one of my production projects.

Before matching anything, measure the data. These are 10 Amazon and 13 Google records from the public Amazon-Google benchmark. Pick a check to see what it flags.

Amazon records

    Google records

      Brute force compares every record with every other. Here that is every Amazon record against every Google record: one square per comparison.

      Now grow the data

      Ten times more records means a hundred times more comparisons. That is the bottleneck to treat before any matching.

      The same bottleneck in one of my production projects, step by step. The figures come from that project; the reduction method itself is not shown.

      1. Data quality controls
      2. Candidate reductionmethod not shown
      3. Detailed matching
      4. Human review of borderline cases

      Request: erase every record of Nora Keller, born 1984-03-12.

      Fictional records: the example from my LinkedIn post.

      Selected work

      • Data quality

        Entity resolution at billion scale

        Finding the records that describe the same company or person across databases with no shared ID. 1.25 billion candidate pairs reduced to about 40,000 comparisons.

        Read the case study
      • Research

        D-HAT: unsupervised matching of heterogeneous records

        An entity matching method that needs no labelled data. 30% higher F1 than the leading unsupervised method, in 1 minute instead of 6 hours.

        Read the case study
      • Privacy and health data

        Linking health records under strict privacy

        US electronic health records matched with insurance claims across millions of records, at 87% accuracy.

        Read the case study
      • Frugal AI

        Choosing the model that earns its cost

        Four open-weight LLMs benchmarked on real data. The one that shipped is 16.5× smaller and 91% cheaper than the first pick, hosted in France.

        Read the case study
      • GenAI in production

        Business analyses written by an LLM, decided by code

        Code computes every indicator and applies the business rules. The model only writes from verified facts, and every change must pass a fixed test set before release.

      • Retrieval

        A document assistant that cites its sources

        Answers questions over a large library of reports and points to the exact page. Every retrieval upgrade was tested one at a time, and only those that earned their cost were kept.

      “We are drowning in information but starved for knowledge.”

      John Naisbitt, quoted in my thesis
      loujain-liekah / about.md
      Background
      15+ years in data: seven years running production databases at CMA CGM, an MSc in machine learning, a PhD in computer science at LIRIS (Université Claude Bernard Lyon 1), then health data, mobility forecasting and generative AI in production.
      Context
      Multilingual and multinational. Studied at universities in France and Germany; worked in a large established group, in startups, in consulting and in academic research with hospital partners.
      Focus
      Entity resolution and data quality, data and privacy engineering for AI, and the technical side of carrying AI from research to production.
      Side projects
      Always a few cooking, in AI and beyond.
      Outside work
      The sea. Fascinated by tides, level 1 diver, and happiest on a nature hike.
      Base
      Lyon, France.
      # about.md
      # Loujain Liekah
      
      ## Background
      15+ years in data: seven years running production databases at CMA CGM, an MSc in machine learning, a PhD in computer science at LIRIS (Université Claude Bernard Lyon 1), then health data, mobility forecasting and generative AI in production.
      
      ## Context
      Multilingual and multinational. Studied at universities in France and Germany; worked in a large established group, in startups, in consulting and in academic research with hospital partners.
      
      ## Focus
      Entity resolution and data quality, data and privacy engineering for AI, and the technical side of carrying AI from research to production.
      
      ## Side projects
      Always a few cooking, in AI and beyond.
      
      ## Outside work
      The sea. Fascinated by tides, level 1 diver, and happiest on a nature hike.
      
      ## Base
      Lyon, France.

      Writing

      Long-form versions of what I post on LinkedIn, on data quality, entity resolution, privacy and AI that is measured before it ships.

      Also on LinkedIn:

      • Five RAG upgrades, measured one at a time
      • A knowledge graph that needs no LLM at query time
      • Not all entities are equal: anchors and hubs
      • How many GPUs does sovereign AI need?

      Read the posts on LinkedIn

      Experience

      1. 2025 to now

        Lead Data Scientist, ADEQUATION

        Generative AI in production, frugal model and infrastructure choices, a document assistant, entity resolution, AI security and causal models. Tech lead of a Bpifrance Pionniers de l’IA programme. Mentoring two junior data scientists and training the whole company on AI.

      2. 2025

        Senior Data Scientist, Fablab Inetum

        Consulting mission: hourly demand forecasting for bike sharing from geospatial and socio-economic data.

      3. 2024

        Senior Data Scientist, Ryte

        Linked US electronic health records with insurance claims across millions of records, at 87% accuracy.

      4. 2020 to 2023

        Researcher, LIRIS, Université Claude Bernard Lyon 1

        Entity matching and clustering of longitudinal patient data, within the Horizon 2020 project QUALITOP.

      5. 2017 to 2019

        MSc Machine Learning and Data Mining, Université Jean Monnet

        Erasmus semester at the University of Freiburg. Research internships at IRIT Toulouse and Laboratoire Hubert Curien.

      6. 2010 to 2017

        Database Administrator and Software Engineer, CMA CGM

        Production databases and the automation of port operations. Managed a team of five system administrators.

      Publications

      Skills

      Leadership
      AI strategy and roadmaps, build-or-buy decisions, hiring and mentoring, communication with executives and business teams
      Data
      Entity resolution, record linkage, data quality, data integration, knowledge graphs, privacy-safe training data
      Generative AI
      LLMs (Mistral, Llama, Qwen), RAG, hybrid search, LLM evaluation, prompt engineering, agents
      Infrastructure
      Frugal and sovereign AI, GPU capacity planning, vLLM, Docker, FastAPI, CI/CD, MLflow, Databricks, Azure, OVHcloud, AWS, GCP
      Security and governance
      Prompt-injection defences, model supply-chain checks, access control, EU AI Act watch
      Machine learning
      Causal inference, survival analysis, forecasting, PyTorch, XGBoost, SHAP, Python, SQL, PySpark