AI Implementation & Assurance: getting AI workflows tested, trusted and adopted in regulated settings · Sydney
For ten years I've delivered software and research operations in settings where mistakes are costly. That includes clinical-trial software rolled out across 20 hospital system instances (56-case UAT; ISO 9001/27001/27017 certification), SOPs for regulated software under FDA 21 CFR Part 11, and data-handling procedures that cut errors from 67% to under 5%. Now I build and evaluate AI workflows myself. Master of Business Analytics, UTS (WAM 81.1).
- HomePreflight · Nebius × NVIDIA Global AI Hackathon
Rule-gated LLM workflow: 20 of 24 cases got the expected urgency tier, against 14 for the model alone, with zero under-triage (the model alone had 4). The red team stopped 49 of 54 attacks and no benign input (0 of 19) was blocked. About US$0.001 per answer, with a spend cap, and a person confirms every extracted line.
Submission frozen for judging until Jan 2027. CI shows red because the public export leaves out the test fixtures used by the development test suite. - FallGuard · 🏆 Best Business Use Case, n8n University Hackathon Sydney
YOLOv8 fall detection feeds an n8n + Claude workflow with human approvals and an audit trail. Cross-dataset testing (recall fell to 2.9% on unseen footage) led me to reframe the product metric as false alarms per camera-day. - KindredRelay · AMD Developer Hackathon: ACT III
Two models benchmarked before their roles were assigned (95.7% vs 86.0% source-grounded answers); human–model agreement κ = 0.797; hazard log, claim-recomputation checks and a published list of what failed. - Taiwan property transactions pipeline · Python · SQL · DuckDB
17 data-quality rules turn 4.83M government sale records into 2.51M analysis-ready residential sales, with a reconciliation funnel and tests.
Python · SQL · DuckDB · Power BI · Claude and OpenAI APIs · n8n · Streamlit · Git · UAT and test planning