Fellow engineer
In depth
Ahmed Maaloul
Ahmed is an AI engineer trained at ESILV, the Léonard de Vinci engineering school in Paris, with a master-level diploma in data and artificial intelligence. He builds assistants that work on a company’s own documents and give answers that can be traced back to their sources.
He created Synapse, an open-source tool that turns documents into a knowledge graph, and has written two short research notes on how to measure whether such a system finds the right passages.
Ahmed builds AI assistants and RAG agents for document workflows on GCP, and works on knowledge graphs and GraphRAG with Neo4j.
He maintains Synapse, an open-source GraphRAG explorer, and audits how retrieval is scored: his two research notes cover a metric with no lower bound, and a reproducible benchmark.
Work so far
- Apr 2026 – today
AI Engineer, Heroiks
Heroiks · Paris
Designs and deploys AI assistants and agents that work on a company’s documents: finding the right passage, writing an answer, and handing it to a person to check.
Looks after each agent from understanding the need to putting it into use, and works on answer quality, user trust and keeping track of costs. Works with technical and business teams to turn AI use cases into working prototypes.
- Role
- AI Software Engineer: assistants and RAG agents for document workflows (ingestion, search, generation, supervision, human validation).
- Build
- Enterprise RAG system on cloud storage, FastAPI, PostgreSQL / pgvector and Vertex AI; backend APIs, ingestion pipelines and LLM orchestration.
- Operate
- Workflows industrialised on GCP with orchestration, observability and containers (Docker, Kubernetes, Argo Workflows). Parsing optimisation, structured logging, cost monitoring.
- Stack
- Python, FastAPI, LangGraph, LlamaIndex, Vertex AI, GCP, pgvector, PostgreSQL, Argo Workflows, Langfuse, Docker, Kubernetes.
- Feb – Aug 2025
AI Engineer Intern, Astek
Astek · Paris
Built a system that matches CVs to job offers, combining fixed rules, similarity search and an AI classifier, with a repeatable way to test how well it works.
Modelled how skills and jobs relate to one another, so the matching understood context and not just keywords.
- Matching
- Hybrid pipeline: rules-based filtering, embeddings and LLM classification, with a reproducible Docker evaluation framework; open-source LLMs benchmarked for classification and explanation.
- Graph
- RDF/OWL ontologies integrated into a Neo4j GraphRAG architecture to enrich context and matching relevance.
- Stack
- Python, Neo4j, RDF/OWL, GraphRAG, Ollama, Pandas, Docker, Linux.
- Sep 2024 – Jan 2025
Data scientist (student consultant), Crédit Mutuel AM
Crédit Mutuel AM, through ESILV · Paris
Combined credit ratings, share prices and market indicators into one dataset, and tested several models to predict rating changes and the risk of default.
- Data
- S&P and Moody’s ratings, stock prices and market indices; class imbalance handled with class weighting and random oversampling.
- Models
- XGBoost, Random Forest and LSTM, with a rolling-window check on temporal robustness (2010–2022). Precision 0.76 on the stable class with XGBoost.
- Stack
- Python, Pandas, XGBoost, LightGBM, LSTM, PCA, financial modelling.
- May – Aug 2024
Mobile Developer, Qovoltis
Qovoltis · Paris
Built a mobile app that lets installers set up electric-vehicle charging stations on their own, cutting set-up time by more than 80%.
- Flows
- Authentication, network pairing, firmware updates and customer sign-off.
- Result
- Configuration time down by more than 80%, less reliance on technical support.
- Stack
- Flutter, Dart, REST APIs, MariaDB, Figma, Postman.
Stack
- Knowledge graph & GenAI
- Neo4j · Cypher · RDF/OWL · ontologies · GraphRAG · RAG · AI agents · LangGraph · LlamaIndex · LangChain · embeddings · vector and full-text search · LLM evaluation
- Programming & data
- Python · FastAPI · Java · SQL · PostgreSQL · pgvector · Pandas · PyTorch · scikit-learn
- Cloud & production
- GCP · Vertex AI · AWS · Azure · Docker · Kubernetes · Terraform · Argo Workflows · GitHub Actions · Langfuse observability
Projects
- Open source
Synapse: GraphRAG knowledge explorer
An open-source tool that reads PDFs, builds a knowledge graph from them, and answers questions with the sources shown. Ahmed created it and maintains it.
- Pipeline
- PDF to knowledge graph: LLM extraction of entities and relations, embeddings, entity resolution, Neo4j storage.
- Retrieval
- Hybrid GraphRAG (vector and full-text), subgraph expansion, sourced answers.
- Delivery
- FastAPI, MCP, CLI, Docker, GitHub Actions and 800+ tests.
- Results
- Hit@1 88%, Recall@8 100%, MRR 0.917.
- Nov 2025
Hackathon Deloitte × Google Cloud
Designed a team of cooperating AI agents that produce marketing material, at a hackathon run by Deloitte and Google Cloud.
- Design
- Multi-agent marketing system built with ADK and Agent Engine; autonomous services on Cloud Run talking over MCP.
- Output
- Marketing assets generated through Vertex AI (Gemini, Imagen).
- Stack
- GCP, Vertex AI, Kubernetes, MCP, Prometheus, Docker, Agent Engine, ADK.
Publications
- Research
A containment metric with no lower bound: auditing retrieval scoring in GraphRAG
Checks a measure used to score document retrieval, shows that it can mislead, and proposes a safer way to score it.
Audit of a containment metric with a null control, replicated on HotpotQA (n=50) and LightRAG; proposes a channel-gated scoring.
- Research
Measuring GraphRAG retrieval in Synapse
A repeatable test bench for judging how well Synapse finds the right passages.
Reproducible benchmark on paraphrased queries: Hit@1 88%, Recall@8 100%, MRR 0.917.
Certifications
- AWSAI Business Strategist, Early Adopter
- AWSCertified Cloud Practitioner
- IT SpecialistJava
- TOEIC870
Languages
- Frenchbilingual
- Englishfluent (TOEIC 870)
- Germanintermediate (B1)
- Arabicnative
Education
- Sep 2023 – Oct 2025
ESILV, Léonard de Vinci engineering school
Engineering diploma (master’s level), major in data and artificial intelligence. Paris.
- Degree
- Diplôme d’ingénieur, major in data and artificial intelligence. GPA 3.88 / 4.00.
- Sep 2020 – May 2023
Polytech International, Tunis
Preparatory classes and the first year of the engineering cycle, major in computer science (software engineering).
- Programme
- Preparatory classes and engineering year 1, software engineering. GPA 4.00 / 4.00.