I build AI systems that ship to production. Not demos platforms.
I own the full stack: data pipelines, model training, multi-agent orchestration, hybrid RAG, secure API design, and MLOps delivery. Currently finishing an Engineering degree in Intelligent Software Engineering at ESPRIT (EUR-ACE, July 2026) while serving as sole architect of DeepCoin-Core at YEBNI a 5-agent LangGraph platform with 47,705-vector hybrid retrieval, EfficientNet-B3 at 80.03% TTA accuracy across 438 classes, and zero hallucination on structured facts delivered solo on a 7-service Docker stack with 122 automated tests and a documented novel ML research finding.
Flagship · DeepCoin-Core
A single coin photograph in. A grounded, cited historical report out. In under 20 seconds.
End-to-end agentic AI platform for archaeological numismatics. Classifies 2,300-year-old coins via CNN, validates with computer vision forensics, and generates hallucination-free historical reports using citation-constrained RAG all through a 5-agent LangGraph state machine with graceful degradation on every path.
┌──────────────────────────────────────────────────────────────┐
│ RAW COIN PHOTOGRAPH │
└──────────────────────────────┬───────────────────────────────┘
│
┌──────────────────────────────▼───────────────────────────────┐
│ Image Preprocessing │
│ HoughCircles autocrop · CLAHE on LAB L-channel │
│ Aspect-preserving resize → zero-pad 299×299 │
└──────────────────────────────┬───────────────────────────────┘
│
┌──────────────────────────────▼───────────────────────────────┐
│ EfficientNet-B3 (80.03% TTA ×8) │
│ 438 classes · ~12M params · 1,536-dim feature vector │
│ Output: class · confidence · top-5 · Grad-CAM++ 19×19 │
└──────────┬──────────────────┬──────────────────┬────────────┘
│ │ │
conf > 85% 40 – 85% conf < 40%
│ │ │
┌──────────▼──────┐ ┌────────▼────────┐ ┌─────▼──────────────┐
│ HISTORIAN │ │ VALIDATOR │ │ INVESTIGATOR │
│ Hybrid RAG │ │ Multi-scale HSV │ │ Vision LLM │
│ BM25 + ChromaDB │ │ 3 crop sizes │ │ (qwen3-vl:4b) │
│ RRF fusion │ │ Ag2S patina fix │ │ KB nearest-neigh. │
│ citation prompt │ │ KB consensus │ │ 9,541-type search │
│ → LLM narrative │ │ override │ │ → cultural matches │
└──────────┬──────┘ └────────┬────────┘ └─────┬──────────────┘
└─────────────────┬┘──────────────────┘
│
┌────────────────────────────▼─────────────────────────────────┐
│ SYNTHESIS AGENT │
│ fpdf2 PDF · Grad-CAM++ heatmap embed · grounded citations │
└────────────────────────────┬─────────────────────────────────┘
│
┌────────────────────────────▼─────────────────────────────────┐
│ FastAPI :8000 · JWT · HSTS · CSP · slowapi · SSE │
└────────────────────────────┬─────────────────────────────────┘
│
┌────────────────────────────▼─────────────────────────────────┐
│ Next.js 15 :3000 · 9 pages · Framer Motion · Zustand │
└──────────────────────────────────────────────────────────────┘
HYBRID RAG ENGINE
BM25Okapi keyword index + ChromaDB cosine search (384-dim)
RRF fusion: score = Σ 1/(60 + rank_r)
→ 5 × [CONTEXT N] citation blocks · 47,705 vectors · < 1ms
LLM FALLBACK CHAIN
1. GitHub Models API (Gemini 2.5 Flash)
2. Google AI Studio (1,500 req/day free)
3. Local Ollama (gemma3:4b / qwen3-vl:4b)
4. Structured KB fallback zero crash, zero hallucination
GRACEFUL DEGRADATION SYSTEM NEVER RETURNS AN EMPTY REPORT
conf > 85% → CNN → RAG → LLM narrative → full PDF
40–85% → CNN → OpenCV validator → KB consensus → PDF
< 40% → Investigator → 3 nearest KB types → PDF
LLM offline → KB fields only → structured report, no hallucination
| Metric | Result |
|---|---|
| CNN accuracy TTA ×8, 438-class | 80.03% |
| Macro F1 438 classes | 0.7763 |
| Knowledge base | 9,541 types · 47,705 vectors (98.2% Corpus Nummorum) |
| Hallucination on structured facts | Zero |
| Test suite | 122 / 122 passing |
| Docker services | 7 |
| End-to-end latency | < 20 s |
| Hybrid search latency | < 1 ms |
| PDF generation | ~0.4–0.5 s |
| Decision | Choice | Why |
|---|---|---|
| CNN backbone | EfficientNet-B3 | Compound scaling fits 4.3 GB VRAM; B7 does not |
| Preprocessing | CLAHE on LAB L-channel | Enhances contrast without destroying metal patina colour |
| Class imbalance | WeightedRandomSampler | 40:1 ratio equalises per-class exposure |
| Regularisation | Mixup α=0.2 + label smoothing 0.1 | Prevents memorisation on 7,677 images |
| Explainability | Grad-CAM++ at features[-4] 19×19 |
3.6× finer than features[-1] |
| Agent framework | LangGraph over CrewAI | Explicit state machine, conditional routing, cycles |
| RAG retrieval | BM25 + ChromaDB + RRF | Keyword + semantic + zero reranker latency |
| LLM grounding | [CONTEXT N] citation blocks |
Structurally prevents hallucination |
| Security | hmac.compare_digest on API keys |
Prevents timing-oracle attacks |
| Thread safety | Double-checked locking on singletons | Prevents OOM races on RAGEngine cold startup |
The same trained model scored 80% on modern photographs but only 15–28% on BNF 1966 catalog scans of identical coin types same model, same classes, different photographic era. An intra-dataset distribution shift not previously documented in numismatic ML literature. Documented in ENGINEERING_JOURNAL.md §184.
| Project | What it does | Stack | Key result |
|---|---|---|---|
| DeepCoin-Core | 5-agent LangGraph platform CNN + hybrid RAG + LLM + full-stack | PyTorch · LangGraph · FastAPI · Next.js 15 · ChromaDB · Docker | 80.03% TTA · 47k vectors · 122/122 tests · zero hallucination |
| Overlord Pipeline | 5-stage multi-LLM video pipeline GPT-4o + Whisper word timestamps + FFmpeg | TypeScript · Node.js · OpenAI · Claude · GROQ · FFmpeg | <15s generation · ~$0.02/video · 100% free-tier capable |
| SkillBridge | AI-augmented MERN skill exchange platform team of 5 | React · Node.js · MongoDB · WebRTC · Socket.io · Gemini API | P2P video · quiz engine · PDF summariser · live collab |
| DevOps Pipeline | 18-stage Jenkins CI/CD quality gates, zero-downtime deploys | Jenkins · Docker · SonarQube · Prometheus · Grafana | 99.98% success rate · < 2 min runtime |
| Voice IoT | Voice-controlled assistive prototype Arduino + Python | Python · Arduino · HC-05 · SpeechRecognition | Hands-free navigation · layered IoT architecture |
AI Engineer PFE · YEBNI, Tunisia · Feb 2026 – Jul 2026
Sole architect of DeepCoin-Core. End-to-end ownership: CNN training (EfficientNet-B3, AMP, Mixup, WeightedRandomSampler, 80.03% TTA), 5-agent LangGraph orchestration, hybrid RAG (BM25 + ChromaDB + RRF, 47,705 vectors), 9-page Next.js 15 frontend, JWT/HSTS/CSP security hardening, 122-test pytest suite, CI/CD on GitHub Actions. Discovered and documented a novel intra-dataset distribution shift in numismatic ML not previously reported in the literature.
Full-Stack Engineering Intern · Tunisia Telecom · June 2025 – Aug 2025
Python automation platform serving 14M+ subscribers. Netmiko across Cisco/Huawei 80% reduction in manual configuration time. Real-time KPI dashboard: < 25 ms latency · 99.5% availability.
Full-Stack Intern · Bright Soft · Jul 2022 – Aug 2022
React/Node.js SaaS features and AI/NLP document processing pipelines.
ESPRIT, Tunis Engineering Degree · Intelligent Software Engineering (EUR-ACE, BAC+5) · 2023–2026
ISIK, Le Kef BSc Computer Science · Software Engineering · 2020–2023
Anthropic Agent Skills · Aviatrix ACE Multi-Cloud · Oracle OCI Associate · AWS Cloud Foundations
Available July 2026 AI Engineer / Full-Stack AI
Open to Relocation
