Most AI requests are a symptom. I work until the real commercial constraint is named — acquisition cost, go-live risk, a process that already fails, or the cost of thinking at scale.
Approach
From commercial constraint to governed, working AI systems.
I help teams turn AI opportunities into systems that can be operated, audited, and improved. Begin with the real business constraint, then define authority, control risk and cost, and deliver a working system.
Enterprises and hiring partners can test the method in the open: name the constraint, set who may decide, and ship a surface that can be audited.
Who may decide, approve, or stop the work is agreed before models are chosen. Authority and spend stay on separate lanes so unsafe decisions and surprise invoices do not travel together.
If a required fact is not in hand, we stop. We do not average our way through a guess. A pause is a successful outcome.
Commercial sends, payments, and go-live stay in draft until a person approves. Software can recommend; it does not quietly commit what cannot be undone.
The deliverable is a working surface — something that can be operated, audited, and improved — not a slide about transformation.
Incomplete data, unverified flags, and unpublished endpoints belong on the surface. Procurement already knows when it is being sold theatre.
Cost hierarchy and fallbacks are design, not an afterthought. Cheap, honest paths first; expensive model calls only where they earn their keep.
Based in Phuket, working remotely and hybrid across APAC. Singapore-based opportunities welcome; engagement structure and work eligibility are discussed case by case.
Method detail / technical appendix
Specialist terms for operators who want the implementation names. They are not required to read the method above.
- Studio authority (who may act) is not InfraDOME spend (which model, at what cost). 7ARCH keeps those routers separate.
- Commander gates cover quotes, secrets, jurisdiction, and Red-compatibility overrides. Missing evidence is HALT, not a guess.
- Reflector go-live checks include RAG quality, white-label, provision dry-run, and sovereignty — ALLOWED or BLOCKED.
- LiteLLM is ACCESS and economics only. Offer and jurisdiction logic do not hide in the spend layer.
- Retrieval uses heuristic-first extraction and RAG-order fallback.
- Production surfaces have shipped on Cloudflare Pages/Workers and Cloud Run, with honest health endpoints.