AM.

I build AI and data systems, and measure whether they hold up.

Language models and retrieval, data pipelines, backend and embedded software. I publish the work here: notes, benchmarks, and research, including the results that did not hold.

Start a project See selected work

What I do

Engineering partnerships for teams building with AI and data.

AI systems in production

Retrieval, language models, and agents engineered for accuracy, latency, and cost. Built to stand up under real traffic, not benchmark demos.

Data and evaluation pipelines

The ingestion, transformation, and evaluation layers your products rely on. Reproducible, observable, designed to outlive their first model.

Software and embedded systems

The backend services, infrastructure, and embedded work that decide whether a model ever ships. Constraints taken seriously, from the data center to the device.

Research

Publications and notes, from the systems I build.

  1. ResearchA containment metric with no lower bound: auditing retrieval scoring in GraphRAGA null retriever that returns only entity names scores 100% fact recall under containment scoring. Why that metric has no floor, and a channel-gated fix that costs honest systems nothing.
  2. BenchmarkMeasuring GraphRAG retrieval in SynapseEight paraphrased queries against a seeded Neo4j fixture: Hit@1 88%, MRR 0.92. What Synapse's eval harness measures, and what it deliberately does not claim.
All research

Projects

Selected projects, built in the open.

01

Synapse: Open-Source GraphRAG Engine

Turns documents into a queryable Neo4j knowledge graph: LLM entity extraction, hybrid vector and full-text retrieval, subgraph expansion, and answers grounded with citations back to the exact source nodes. Ten swappable LLM providers, cloud or fully local, behind one environment variable.

  • Hit@1 88%, MRR 0.92 on its retrieval eval
  • 800+ tests, CI, integration suite
  • Runs in 60s with Docker, no API key
  • Dual-licensed: AGPL-3.0 / commercial
Source
GraphRAG Neo4j LangChain FastAPI Next.js
Synapse: Open-Source GraphRAG Engine interface
02

Insurance Review Rating Predictor

Predicts customer satisfaction scores from raw insurance reviews. Benchmarks TF-IDF and neural baselines against RoBERTa and a LoRA-fine-tuned LLaMA 3.2, with SHAP attributions explaining each prediction. Served as a public Streamlit app.

NLP LoRA RoBERTa SHAP
03

Yogurt PLM Platform

Product Lifecycle Management platform built for a dairy producer: bill-of-materials and recipe tracking, supplier and customer records, SolidWorks file integration, invoicing, and per-role dashboards on a MongoDB backend.

PLM Full-stack MongoDB RBAC

Earlier work

  • EcoStay: Sustainable Hotel Recommendation

    Geospatial and semantic hotel recommender for Paris: RoBERTa embeddings ranked with Haversine distance, served with FastAPI in Docker.

  • Pothole Detection with YOLO & DETR

    Benchmark of YOLOv5–v11 and DETR on custom pothole datasets with Mosaic and MixUp augmentation; results tracked in Tableau.

What people say

A few words from people I've worked with.

He took ownership of complex tasks and delivered with both rigor and creativity, a rare blend of research-oriented mindset and engineering discipline.
Wafaa El Husseini

PhD · Data Scientist & AI Engineer · Astek

I taught Ahmed during his preparatory cycle at Ecole Polytechnique Internationale: disciplined, autonomous, and gifted in problem-solving.
Marouane Ben Haj Ayech

Computer Science Teacher · Polytech Intl

Ahmed is a talented developer with solid skills across security, development, and networking, which he puts to use effectively to advance whatever project is entrusted to him.
Pierre Lemère

iOS Developer · Swift & SwiftUI · Qovoltis

About

I work in the gap between a model that demos well and a system that earns its keep in production.

I'm an engineer trained at ESILV (Data & AI, M.Eng. equivalent), based in Paris. My work sits wherever AI and data meet real constraints: language models and retrieval, the pipelines that keep them accurate, and the software and embedded systems they run on.

I've shipped LLM evaluation systems and predictive models inside an asset manager, an engineering consultancy, and an EV-charging product. Across those, the same pattern: define the outcome clearly, choose the smallest system that delivers it, and instrument it so nothing degrades quietly.

Ahmed Maaloul
Location
Paris, France
Languages
  • Arabic Native
  • French Bilingual
  • English Professional
  • German Intermediate

Experience

Where I've worked and what I shipped.

  1. Apr 2026 – Present
    Paris

    AI Engineer · Heroiks

    Building AI agents and RAG / GraphRAG pipelines end-to-end, from prompt architecture and evaluation through to production deployment.

    Python LLMs RAG GraphRAG Agents
  2. Feb – Aug 2025
    Paris

    AI Engineer · Astek

    Designed and evaluated an LLM-driven CV–job matching system across DeepSeek, Gemma, LLaMA, Mistral and Phi. Built the inference and evaluation pipelines, and co-authored a research paper on hybrid explainable matching (defended 18.5/20).

    Python LLMs RAG Docker Evaluation
  3. Sep 2024 – Jan 2025
    Paris

    Data Scientist · Crédit Mutuel Asset Management

    Built predictive models for corporate-issuer rating changes and default probabilities, with temporal feature engineering and class-imbalance handling. Delivered an explainable AI tool (SHAP/LIME) used by analysts on simulated investment pipelines from 2009 to 2024.

    Python XGBoost LSTM LightGBM SHAP
  4. May – Aug 2024
    Paris

    Mobile Developer · Qovoltis

    Shipped a cross-platform Flutter app for configuring EV charging stations: QR scanning, hotspot pairing, multilingual UI. Reduced hotline calls by 80%, ran hardware testing across embedded and UI teams.

    Flutter Dart IoT

Stack

Tools I reach for, chosen for the job, not for the resume.

AI & ML

Python PyTorch scikit-learn Transformers RAG Ollama SHAP

Data & MLOps

FastAPI Docker Airflow Spark Postgres BigQuery GitHub Actions

Backend

Node.js Express MongoDB PostgreSQL

Frontend

React TypeScript Tailwind Vite

Education

Engineering training, Data & AI specialization.

  1. 2023 – 2025
    Paris

    ESILV

    Diplôme d'Ingénieur (M.Eng. equivalent) · Data & AI

    Internship defense 18.5/20

  2. 2020 – 2023
    Tunis

    Polytech Intl

    Integrated Preparatory Cycle, Engineering Programme · Computer Science

Get in touch

Let's build something that earns its keep.

I take on a small number of engagements at a time. If you're shipping AI features and want a thoughtful technical partner, send a note. A few sentences on the problem is all I need to tell whether I can help.