Nduka Nwagbo
AI Engineer · Birmingham, UK
Profile
I turn open-weight models and open-source tools into AI that works in any industry: running where your data lives, grounded in your own documents, and tested before anyone relies on it.
I build AI that professionals can check before they trust it, and I take it through the whole lifecycle: data, modelling, validation, explanation, deployment and handover. I test models against real-world data, not only a test set, and I write down what each system cannot do.
Before AI, I studied communication and multimedia design and worked in IT support, explaining fixes to people who are not technical. I still write for the people who will use the system.
Selected projects
US accident impact risk
Found that the popular “accident severity” label mostly tracked which traffic feed filed the report, then rebuilt an honest model on 3.2 million accidents from one consistent feed, explained it, and made it run entirely in the browser.
Result: 0.852 ROC-AUC on later, unseen months, with no server
us-accident-impact-risk.nduka-nwagbo.workers.dev github.com/Nduka99/accident_severity_xai
Consumer complaint triage
Classifies US consumer financial complaints into ten products with distilled transformers, finds the relevant CFPB examination procedures, and decides whether each case is handled automatically or sent to a person.
Result: Macro-F1 0.757 on 274,065 held-out complaints, up from 0.689
complaint-triage.nduka-nwagbo.workers.dev github.com/Nduka99/customer_complaint_triage
PetroKnowledge
Answers questions about public NNPC information with local open-weight models. Every claim carries a quote that code checks word for word in its source, figures are checked again by a second model, and unsupported questions get a clear refusal.
Result: Runs on local open-weight models; every quote is checked against its source
PPE compliance detection
Merged three public datasets into 10,864 images, trained YOLO11 and served it as an app that flags missing hard hats and vests. An independent real-world benchmark exposed a model that missed real vests; two fixes raised real-world vest recall from 0% to 50.5%.
Result: Real-world vest recall raised from 0% to 50.5% at 80% precision
ppe-detection-blush.vercel.app github.com/Nduka99/ppe_compliance_detection
Agentic menu development
My MSc dissertation. A five-stage agentic workflow turns a hospitality client's intake form into a consultant-grade menu brief. Each stage reads only what the approved stage before it hands over. Open-weight models draft, another model family reviews, code checks what they miss, and a consultant approves each step.
Result: Judged usable with minor edits by a practising consultant
SpineLens AI
The evidence toolkit behind The Innovation Spine, a team proposal to reconnect Birmingham's Knowledge Quarter to the city centre. Route legibility scoring, wayfinder placement, traffic and collision evidence and a costed plan, shipped as an interactive map.
Result: The route ranking held in 86% of 4,000 random weightings
ApexMind
A four-hour build for the BCU AI Hackathon 2026. Answer 100 multiple-choice questions using only open models of 8 billion parameters or fewer, grounded in Wikipedia, with confidence scores and repeatable runs.
Result: 100% on a hand-checked 20-question set, up from 90% without retrieval
QHybrid Lab
A web service that runs a quantum-kernel classifier against classical models on the same rows, features and tuning folds, for a cybersecurity and a medical dataset, and serves a calibrated network-intrusion triage model. Every comparison can be re-run from the public code.
Result: With a tuned bandwidth the quantum kernel tracked the RBF SVM, at or below the best classical model in 31 of 40 runs
Experience
Recruitment Ambassador
Explain technical programmes to prospective students and families at open days and campus tours.
IT Support Technician
Kept systems reliable for non-technical staff and explained fixes clearly to users and management.
Education
MSc Artificial Intelligence
Distinction (formal award pending). Dissertation paper 85, viva 81. Modules in applied AI, natural language processing, computer vision, advanced machine learning, and AI ethics, bias and governance.
BSc (Hons) Communication and Multimedia Design
Recognition
- UKREiiF Next-Gen Challenge, West Midlands winners, 2026. With my team. Our Knowledge Quarter proposal was presented at Birmingham City Council.
- BCU AI Hackathon, winners, 2026. With my team.
- PGXPO, Silver award, 2026. For my research poster.
Certifications
- CompTIA Security+ (SY0-701) and Network+ (N10-009), valid to 2027
- Certificate in Cybersecurity, University of Windsor, 2025
- Linux Foundation LFS151 Cloud Infrastructure, LFS158 Kubernetes and LFS250 Kubernetes and Cloud Native Essentials, 2025
Skills
- Machine learning
- PyTorch, Transformers, scikit-learn, LightGBM, XGBoost, CatBoost, YOLO11, knowledge distillation
- LLMs and retrieval
- local open models with llama.cpp, GGUF quantisation, agentic workflows with human approval, LLM review across model families, RAG, BM25, pgvector, cross-encoder reranking, schema-constrained generation, document extraction
- Evaluation and explainability
- SHAP, Integrated Gradients, time-based splits, bootstrap intervals, McNemar's test, calibration, independent benchmarks, evaluation harnesses, pre-registered evaluation
- Software and deployment
- Python, TypeScript, SQL, FastAPI, React, Postgres, Docker, GitHub Actions, Cloudflare, Hugging Face Spaces, AWS, Terraform, Playwright, pytest