Anas
Makrane

Data Scientist & AI Engineer

Engineering intelligent solutions from the ground up. I design the underlying data infrastructure, train the predictive models, and evaluate the language systems that turn complex information into clear insights.

fig.00 · everything downstream of data

A graph centred on data. Data connects to 6 domains:machine learning, llms, knowledge graphs, pipelines, reports & bi, quality. Each domain connects in turn to the tools and techniques listed under it:machine learning — prediction, feature eng., shap; llms — rag, graphrag, agents; knowledge graphs — neo4j, ontology; pipelines — airflow, spark, terraform; reports & bi — power bi, streamlit; quality — evaluation, traceability.

The Problem

Intelligent systems fail when their data foundation cracks. These failures don't happen overnight—they compound silently until the entire architecture becomes unreliable.

  1. [01]

    Unverifiable Outputs

    Plausible answers are dangerous if they can't be traced. Without grounded retrieval, the line between fact and fabrication blurs, exposing the business to unseen risks.

  2. [02]

    Siloed Unstructured Data

    High-value domain knowledge remains locked inside static documents and isolated formats, forcing teams to rely on fragmented, manual discovery.

  3. [03]

    Opaque Decision Making

    Predictive models that cannot surface their driving factors fail to earn user trust, leading to poor adoption and wasted engineering effort.

  4. [04]

    No baseline, so no progress

    Iterating without strict quantitative baselines turns deployment into a guessing game. Without a framework to measure quality, you aren't engineering; you are just experimenting.

Most teams try to patch these architectural flaws with larger LLMs. But structural problems require structural solutions: transparent data lineage, logically sound graph schemas, and deterministic evaluation frameworks.

Featured Work

Built for myself, start to finish. Both are on GitHub.

01 / Cloud Data Platform & NLP2025

Cloud-Native Game Intelligence Platform

An easily deployable, production-ready data platform designed to extract player sentiment, playtime correlations, and community trends from unstructured Steam reviews. Built from the ground up on Azure with Infrastructure as Code to guarantee reproducible deployment and transparent data lineage.

  • 230,000+Reviews processed
  • Python
  • Terraform
  • Azure
  • Docker
(opens in a new tab)
02 / Deep Learning & Audio DSP2024

Acoustic Spectrogram Classification Engine

A research-driven deep learning pipeline that transforms raw audio into mel-scaled spectrograms to evaluate spatio-temporal feature extraction across musical genres. Rigorously benchmarked across four distinct architectures and varying temporal slice windows to achieve optimal inference speed and accuracy.

  • 94%Accuracy
  • 4Architectures compared
  • Python
  • PyTorch
  • Librosa
  • Scikit-Learn
  • HuggingFace
(opens in a new tab)

In Production

Systems running for other people, each one measured. The numbers are the ones I could verify, not the ones that sounded best.

DARYLMar 2026 — present

Turning client documentation into a queryable graph

AI Engineer

Designed a modular GraphRAG pipeline converting unstructured technical documentation into a queryable Neo4j knowledge graph. Delivered two client POCs using an architecture that guarantees traceability and mathematically prevents fabricated citations.

  • +57%Extraction quality
  • 0%Invented sources
  • 96%Search relevance
  • Python
  • Neo4j
  • GraphRAG
  • Azure OpenAI
  • LangChain
  • LangSmith
  • OCR
  • Docker
Attijariwafa Bank — Trade FinanceFeb 2025 — Aug 2025

Pricing 2.7 million trade finance operations

Machine Learning Engineer

Modeled pricing across five years of history to maximize portfolio profitability. Evaluated 16 algorithms to reduce the prediction error to 4.45%, and shipped an explainable scenario simulator directly adopted by the commercial team.

  • 2.7MOperations modelled
  • 16Algorithms compared
  • 4.45%Mean error
  • Python
  • scikit-learn
  • Pandas
  • NumPy
  • SHAP
  • Streamlit
Département de l'ArtisanatJun 2024 — Aug 2024

A RAG assistant for Moroccan craft expertise

AI Intern

Built an end-to-end retrieval-augmented chatbot to answer specialist questions and preserve institutional knowledge. Engineered the complete vector search architecture to optimize the retrieval of contextualized answers.

  • 13+Crafts covered
  • 100%End-To-End pipeline
  • Python
  • LangChain
  • Pinecone

Got data nobody can answer questions about?

Reach out ↓