Final-year engineering student at ESPRIT (Software & Data Science). I build LLM systems the way I'd want to review them:
every claim comes with a number, and every number comes with a test.
$ release-gate assess --candidate aymen-rahmani --target pfe-2027
+ retrieval Recall@5 0.73 to 0.94 after retrieval tuning (Azure_RAG, 40+ eval runs)
+ answer quality 0.93 correctness vs 0.81 baseline, 60 questions (Azure_RAG, agentic RAG)
+ safety 0 of 240 prompt-injection attacks succeeded (Azure_RAG)
+ regression CI eval gate blocks any PR that lowers quality (Azure_RAG, FuelCell-RAG)
+ grounding 0 invalid citations, 7/7 correct refusals (FuelCell-RAG, DE/EN)
- open item faithfulness 0.87 vs 0.90 target, calibrating (FuelCell-RAG)
verdict: ALLOW | available February 2027 | 4-6 monthsYes, the red line stays in. A gate that never says no isn't a gate.
| Project | What it proves | |
|---|---|---|
| ☁️ | Azure_RAG EU regulatory assistant · GDPR, AI Act, NIS2, DORA |
Hybrid RAG, GraphRAG and agentic RAG behind one streaming API on Azure, deployed with Terraform and GitHub Actions; versioned eval harness, CI eval gate, PII masking, OpenTelemetry tracing, 139 tests |
| 📚 | FuelCell-RAG bilingual technical-document RAG · in progress |
German/English retrieval over datasheets, requirement specs and STEP files; every answer cites a page, table cell, requirement ID or CAD part; Recall@5 0.925 on a 60-question golden set |
| 🌊 | CrisisLink-5DS multimodal flood triage · team of 6 |
Telegram bot (photo, voice in Derja/French/English, GPS) on n8n; vision and distress agents, a P1–P4 triage agent I own, multilingual RAG on Qdrant, deterministic safety guardrail, human escalation |
| 🛡️ | Security & Quality release gate my agent in a 6-agent DevOps/MLOps platform |
Tests, coverage, lint, CVEs (Trivy), secrets (Gitleaks) and SonarQube merged into one ALLOW / WARN / BLOCK verdict, wired into CI and exposed as an MCP server |
flowchart LR
D[Docs & data] --> R[Retrieve<br/>hybrid + rerank]
R --> A[Reason<br/>agents + tools]
A --> E{Evaluate<br/>golden set · judge · attacks}
E -- regression --> R
E -- pass --> G[Gate<br/>CI · ALLOW / BLOCK]
G --> S[Ship<br/>traced · costed · reversible]
LLM & agents: LangChain · LangGraph · MCP · hybrid RAG · GraphRAG · agentic RAG · tool calling · n8n · Ollama
Eval & safety: Recall@k · MRR · faithfulness · LLM-as-a-judge · prompt-injection testing · PII masking · guardrails
Data & ops: Qdrant · ChromaDB · Azure AI Search · OpenTelemetry · Langfuse · MLflow · Kafka
Arabic (native) · English (C1, certified) · French (B2) · German (B1–B2, TestDaF)
Currently: finishing FuelCell-RAG's faithfulness calibration · open to a PFE internship in Europe from February 2027