Conversational AI assistant that lets ~80 maintenance engineers query repair procedures, fault codes, and task guidance from aviation technical manuals — replacing hours of manual document search.

Architecture: RAG retrieval backbone (Qdrant) + Agent orchestration layer + Claude Vision + GPT-4o
Status: ✅ MVP delivered · 🔄 Live operational testing with structured bug-capture pipeline
My role: Product owner — defined scope, designed architecture end-to-end, structured validation framework

RAG Agentic AI Claude Vision GPT-4o Qdrant Flask AWS Bedrock

AI system that auto-generates Component Maintenance Manual (CMM) drafts from engineering source documents — reducing first-draft preparation time by ~65%.

Deployed to: ✅ Safran Electrical & Power · ✅ Safran Seats · 🔄 Safran Landing Systems (in progress)
Stack: GPT-4o + Claude via AWS Bedrock · PyMuPDF + Pandas extraction pipeline · Structured output pipelines
My role: Sole architect and delivery lead — requirements through multi-BU production rollout

GPT-4o AWS Bedrock Claude PyMuPDF Pandas Prompt Engineering Docker

Multimodal RAG platform indexing 1,000+ aviation PDFs, serving ~50 engineers with sub-2s query response — reducing documentation search time by ~70%.

Stack: Claude Vision + GPT-4o + Qdrant · Flask REST API (15+ endpoints) · Async processing · Hybrid AWS S3 + local cache · Docker · GitLab CI/CD
Roadmap: Multi-agent GraphRAG expansion · Specialised domain agents

Qdrant LangChain Claude Vision GPT-4o Flask Docker AWS S3 GitLab CI/CD

Inherited a failed AI initiative (abandoned random forest system). Diagnosed root-cause failures in model architecture, data transformation, and class imbalance handling — then redesigned the solution from the ground up.

What changed: Pure ML approach → Hybrid architecture: lightweight binary triage classifier + specialised second-stage model + structured DB classification layer
Result: ~38% accuracy improvement over predecessor · Phase 1 delivered on schedule · Phase 2 in progress
Key call: Convinced stakeholders that ML alone was the wrong tool — the most important product decision on the project

Scikit-learn Class Imbalance Split-model Architecture Feature Engineering Pandas Python

End-to-end computer vision system for real-time overhead bin space utilisation assessment on aircraft — from TRL 1 (concept) to TRL 4 (demo-ready prototype).

Stack: YOLOv8 + MaskRCNN + UNet · Single-camera geometric 3D reconstruction · NumPy spatial computation
Performance: 95%+ detection accuracy · <100ms latency via OpenVINO on Intel NCS2/Jetson Nano
Exhibited at: AIX Hamburg Airshow

YOLOv8 MaskRCNN OpenVINO Intel NCS2 Jetson Nano 3D Reconstruction Edge AI

Phase 1: Stereo vision food inventory pipeline at 30–50 FPS. Phase 2: Multi-camera 3D reconstruction with Kalman filter + Hungarian algorithm multi-object tracking for galley-wide monitoring.

Partnerships: Luxonis (hardware) · Roboflow (annotation) · Passio Life (nutritional data)
Exhibited at: AIX Hamburg Airshow

OAK-D Intel Myriad X YOLO Kalman Filter Hungarian Algorithm OCR Multi-Camera

Comparative study of RAG versus PEFT/QLoRA fine-tuning (Mistral-7B-Instruct-v0.2, 4-bit NF4 quantization) for domain-specific physics QA.

Evaluation: LLM-as-a-Judge harness across 4 conditions — Base · Base+RAG · Finetuned · Finetuned+RAG
Stack: LangChain · FAISS · SentenceTransformers (all-MiniLM-L6-v2) · TRL/PEFT · RunPod GPU

Mistral-7B QLoRA PEFT LangChain FAISS LLM-as-a-Judge RunPod