Applied AI systems builder | Evidence-bound workflow controls
I build and test small AI systems for work where a convincing answer is not enough. The focus is clear inputs, limited authority, repeatable checks, and evidence a human can inspect.
| Repository | What it demonstrates |
|---|---|
| ASPRON | A browser-based Safe Intake Capsule: risky fields are identified, raw access is blocked, a human approves a redacted copy, and reduced evidence is retained. |
| AI Policy & Terms Analyzer | Python tooling that extracts technical and operational signals from public policies, terms, and privacy pages for human review and structured output. |
| KinSync | A full-stack family-coordination application scaffold with separated web, API, auth, billing, database, and email concerns. |
- Establish what the available evidence supports, then state what remains unknown.
- Keep model output, policy, verification, and human authority distinct.
- Build the smallest testable system that can prove or disprove a claim.
- Treat documentation, tests, and reproducible runs as part of the deliverable.
- Keep prototypes honest about their boundaries and production status.
- Safe AI intake, policy gates, and reduced evidence receipts
- AI workflow evaluation, human review, and control design
- Source-grounded document analysis and practical full-stack systems
A useful AI system should show what it used, what it did, what it could not establish, and who remains authorised to decide.