Requirements you can trust. Delivery you can stand behind.

Elicited, documented, and validated business requirements across a live digital transformation programme, producing BPMN 2.0 process maps and functional specifications.

• Facilitated workshops with 8 stakeholders, three times weekly, translating business needs into clear, actionable requirements.
• Designed and executed 10 manual test cases against real-world and edge-case scenarios; achieved 100% accuracy against specification, zero defects to production.
• Evaluated and documented failure modes each sprint to support evidence-based release readiness decisions.

A full Business Analyst delivery cycle run independently, without a brief or assigned scope: discovery, current-state process mapping, opportunity identification, business case production, BPMN process design, RACI governance, benefits realisation, change management, governance documentation, and AI agent design, applied to a fictional social care organisation using real operational knowledge from the care sector.

The problem

Support workers in residential care homes lose an estimated 40–72 hours per week across a three-shift team to manual administrative logging, 5–15 minutes of every hour spent on devices instead of care. Medication records carry the highest error risk, and early health deterioration is often caught only through informal verbal observation, lost between shifts.

Four opportunities identified and quantified

Combined annual value: £100,000–£160,000 per home.

Grounded in original research, not assumption

Every recommendation traces back to my MSc dissertation on federated learning for threat detection, the same anomaly-detection mechanism (train on normal, flag high reconstruction error) transfers from network intrusion detection to care log and health-baseline anomaly detection. Dissertation results: 0.95 accuracy, 0.95 F1-score, 0.99 AUC, and 0.9934 SHAP stability under bootstrapped perturbation, with performance held under 30% feature corruption.

Governance built in, not bolted on

• Full AI Governance Pack: Acceptable Use Policy, six-risk Risk Register with likelihood/impact/mitigation/owner, and a CQC-auditable AI Decision Log template.
• Federated architecture throughout, individual health data never leaves the care home server, satisfying UK GDPR special-category handling.
• Human sign-off mandatory before any AI output is used; SHAP-explained audit trail for every flagged decision.
• Benefits Realisation Framework defining baseline, target, measurement method and named owner for every opportunity.
• Change & Adoption Pack: plain-English colleague communications, a shift-floor prompt guide, and a four-week Early Adopter champion framework.

KB Matrix connects clients with Afro and textured hair stylists across the UK and Europe. I test client and stylist-facing flows ahead of production release.

• Blocking defect: Stylist email/password sign-up returned a "Bad Request" error, preventing new stylist onboarding entirely.
• UX defect: Saving a client profile preference didn't return the user to their profile, trapping them behind a non-dismissible confirmation popup.
• Platform inconsistency: Notifications settings skipped the native permission prompt and redirected straight to system settings, breaking standard iOS convention.

A service evaluating AI systems for hallucination, prompt injection, jailbreak, and policy failure risk, mapped to established frameworks rather than ad hoc checks.

• Exploring bug bounty and red-teaming routes (Mozilla 0din, Gray Swan Arena) as proof-of-skill before first paid client.
• Underpinned by original federated learning research from my MSc dissertation on anomaly detection and SHAP explainability under adversarial conditions.

A federated learning pipeline for Zero Trust intrusion detection, comparing centralised and federated training on the UNSW-NB15 dataset, the empirical question underneath every audit I run: can a system stay accurate without centralising sensitive data.

• Federated model reached 0.95 accuracy, 0.95 F1-score, 0.99 AUC, matching centralised performance while keeping all training data on local clients.
• SHAP explanations held 0.9934 Spearman rank stability under bootstrapped perturbation, and performance held under 30% feature corruption.
• Includes an honest scoping note on what was cut from the original three-domain proposal, and why, plus a documented failed approach (hybrid feature selection collapsing to 0.14 accuracy) kept in the write-up because the failure was informative.

Reserved for the first applied audit, a real evaluation of a public AI system, written up with the same rigor as the case studies above. No placeholder claims here until there's a real one to show.