I build governed AI. Not prompts. Architectures. Systems that are not allowed to fail.
Principal Engineer · Agentic AI · Core Banking (Temenos T24) · OWL2 · MCP · RAG/KAG · World Bank & Central-Bank Programs
As a Principal Engineer working across core banking (Temenos T24) and World Bank & central-bank programs, I sit where data science meets enterprise architecture. I don't just train models — I architect the governed systems that put them into production where failure is not an option. Years shipping fiscal & banking infrastructure — a fiscal POS certified by the Ministry of Finance across 5,000+ stations, and an ODS certified on Temenos Exchange — taught me what "production-grade" actually costs. I bring that same discipline to AI: observability, idempotency, durability, and governance, applied to LLMs and semantic reasoning.
My edge: most people can train a model or ship a system. I do both — and I make the model behave inside the system.
I build agentic, knowledge-grounded AI that operates inside a governed semantic world — not a chatbot, an architecture:
flowchart LR
D["Domain Data<br/>SQL · Documents · Streams"] --> S
subgraph GOV["🔒 Governed Semantic Layer"]
direction TB
S["OWL Ontologies<br/>SWRL Rules · RDF4J"] --> K["KAG Retrieval<br/>SPARQL · Vector · BM25"]
end
K --> A
subgraph AGENT["🤖 Agentic Orchestration"]
direction TB
A["LLM Reasoner<br/>think · act · observe"] <--> T["Tools · MCP<br/>code · search · gen"]
A --> V["Verify · Guardrails<br/>adversarial checks"]
end
V --> O["⚙️ Governed Action<br/>sign · persist · serve"]
O --> M["📈 Observability<br/>cost · durability · audit"]
M -. feedback .-> A
| Capability | What I architect |
|---|---|
| Agentic orchestration | Multi-agent think/act/observe loops, tool-use, MCP servers, parallel dispatch, self-correction bounds |
| KAG — Knowledge-Augmented Generation | LLMs reasoning inside OWL/SWRL/RDF4J ontologies — contextual retrieval (SPARQL + vector + BM25), not naive RAG |
| LLMOps | Multi-provider routing, prefix/KV caching, cost metering, streaming, durable task recovery |
| Model serving | PMML / JPMML portability — train in Python, serve in Java at enterprise scale |
| Governance & trust | Adversarial verification, guardrails, tamper-proof audit trails, XAdES/PKCS#11 digital signatures |
| Production discipline | Idempotency anchors, single-source-of-truth state, observability, zero-failure SLAs |
A reactive Operational Data Store for Temenos T24 core banking — certified and published on Temenos Exchange (the marketplace of 3,000+ member banks) and adopted by 3+ banks (ATB, Baraka Bank, BH). @ UniQ Soft Technology.
What it is. A signals-based, low-latency ODS that models the data-warehouse architecture (COB export via T24 DW.Export) and drives a compliance-grade account-reconciliation workflow engine with a full audit trail.
More than a data store — what I built around it:
- Semantic AI layer — OWL ontologies (PMBOK · ISO 31000 / 27000 · IFRS · FRM) + SWRL inference + RDF4J/SPARQL agents reasoning over live banking data.
- Offshore Risk KPIs for Warba Bank (Kuwait) — Market / Liquidity / Cost risk via SSIS ETL from T24 FRM → SSRS, with zero COB breach.
- End-to-end MLOps (CRISP-DM) — predictive & classification models on internal banking data, served via PMML.
Why it matters. Core-banking data is unforgiving — a COB breach is a regulatory event, not a bug. This shipped certified on Temenos Exchange, with zero COB breach.
Temenos T24 · Java / JEE · Django · RDF4J · OWL · SWRL · SPARQL · SSIS / SSRS · Oracle PL/SQL · Kafka · Redis · gRPC · Spark / Scala | 🔒 Described faithfully from the record.
Hands-on, end-to-end — from raw signal to served decision:
- Computer Vision / OCR — Arabic document OCR pipelines (deskew sweeps, glare/label removal, multi-engine fallback) on real Tunisian ID & fiscal documents.
- NLP / NLU — NER + fuzzy entity resolution, semantic invoice checkers, intent classification across EN / FR / Tunisian.
- Classical ML — risk scoring (credit default, tax-risk), feature selection (RFE/RFECV), dimensionality reduction, cross-validated model selection.
- Semantic AI — ontology-driven fraud detection with SPARQL + SHACL over knowledge graphs.
📊 Deep-dive ML / DS portfolio → @MuhamedHabib
No placeholder projects — everything below runs in production.
| Project | What it is | Stack |
|---|---|---|
| ODS — Temenos Exchange (T24) | Reactive Operational Data Store certified & published on Temenos Exchange (3,000+ member banks); adopted by 3+ banks (ATB, Baraka, BH). T24 DW.Export + reconciliation engine + semantic AI layer. | Temenos T24 · Java · RDF4J/OWL · Oracle |
| T24 Risk KPIs — Warba Bank (Kuwait) | Market / Liquidity / Cost risk KPIs via SSIS ETL from T24 FRM → SSRS — zero COB breach. | Temenos T24 FRM · SSIS · SSRS |
| PMIS Madagascar — World Bank | Full-stack government platform, Ministry of Energy & Hydrocarbons — conception → production. | Java 21 · Spring Boot 3 · Angular 17 · Spring Batch · Docker |
| Fiscal POS — Ministry-Homologated | Cash-register system certified by the Ministry of Finance, 5,000+ stations, tamper-proof audit trail, zero critical failure. | XAdES · PKCS#11 · Remote Agent/Client |
| Fatoora Hub (active) | End-to-end El Fatoura e-invoicing: draft → sign → submit → accept. | XAdES · TunTrust · Spring Boot 3 |
| ✦ MCP Orchestration Research (active) | Multi-agent pipelines where LLMs operate inside governed semantic worlds (OWL + SWRL + RDF4J). | Claude AI · MCP · KAG · SPARQL |
AI · ML · Data
Agentic & Semantic AI
Enterprise Backend & Platform
Cloud & DevOps
🏆 Pair Extraordinaire ×3 · Pull Shark ×3 · YOLO · Quickdraw | 🌍 GSoC 2026 — Accord Project (Linux Foundation): agentic workflow + LLM template-logic executor.
I worked before AI — and with AI. That dual lens is my speed and my depth. I don't arrive alone at a mission; I arrive with an amplification capability — orchestrated agentic systems that deliver in one day what a team handles in a week, without trading away the governance, audit, and zero-failure discipline an enterprise demands.
Available — Remote · On-site · Relocation (France)