Senior Automotive Software Engineer — bridging safety-critical embedded systems and graph data science.
11+ years building automotive software at Valeo, TTTech Auto, and Concentrio AG. Currently applying Neo4j Graph Data Science and Python to automotive software validation at scale.
🌐 mohamed-soubhi.github.io · 💼 LinkedIn · 📊 Kaggle
- AUTOSAR embedded software (BSW / RTE / ASW) on ARM Cortex-M7 and AURIX TC3xx
- Functional Safety — ASIL B, ISO 26262, ASPICE Level 4
- Graph-based software validation — Neo4j signal-flow analysis for automotive code quality
- Data pipelines for automotive software analysis (Python, Pandas, Docker)
- Diagnostic software — UDS / ISO 14229-1, CANalyzer, Davinci Configurator
Embedded
Data Science
Tools
End-to-end fraud detection knowledge graph on PaySim synthetic transactions (50k rows).
Neo4j 5 + GDS 2.13 · GraphSAGE GNN · LangChain NL→Cypher · Ollama Cloud (deepseek-v4-flash)
GDS Algorithm Pipeline — Account→Account virtual graph projection:
| Node Property | Algorithm | Fraud Signal |
|---|---|---|
community |
Louvain | High-fraud-density clusters |
pageRank |
PageRank | Central money-hub accounts |
wccComponent |
WCC | Isolated fraud rings |
betweenness |
Betweenness Centrality | Bridge / relay accounts |
triangleCount |
Cycle Detection (Cypher) | Circular layering flows (A→B→C→A) |
GNN Layer — 3-layer GraphSAGE trained on GDS properties as node features
→ writes fraudProb ∈ [0,1] to every account · ensemble with rules maximises recall
Fraud Rules — 3 Cypher pattern queries: velocity (>3 txns in 10 steps) · mule chain (A→B→C→cashout) · balance drain (≥95% emptied)
Benchmark — WCC 20ms · sampled Betweenness 142× faster than exact · PageRank converges in 2 iterations