Oracle FM website
A corporate site for a contractor whose clients care about compliance and building safety, which needed to be editable without code and to answer practical questions quickly.
Direct client. Oracle FM (Oracle FM Ltd and Oracle Managed Services LLP), a UK facilities-management and building-safety contractor, briefed him directly; he reports to them directly and maintains the site. From early 2026; ongoing.
His role
- DocumentedSole developer: every commit in the repository is his (82, January to October 2026).
- DocumentedBeyond code: discovery and content model, a style guide, migration from the previous WordPress site, analytics and Search Console reporting, SEO audits and client reporting.
His contribution
- DocumentedContent as data with GitHub as the CMS: editors use a CMS interface that writes through the GitHub API as branches and pull requests, so developers and editors share one source of truth and one history; content is validated against Zod schemas before it is saved.
- DocumentedA site assistant on Google's Gemini, grounded in the site's own content.
- DocumentedA single omnibox for search, questions, tools and actions. Intent routing is deterministic first (rules and an alias map); only requests that match nothing go to the model.
- DocumentedFifteen facilities-management calculators (fire-risk estimate, alarm sizing, maintenance cost and similar), implemented as deterministic rules, not by the model.
- DocumentedForms on Netlify's Postgres database and email through Resend.
Decisions and trade-offs
- DocumentedCalculations stay deterministic and the model answers questions; the omnibox decides which is which before any model is called.
What was delivered
- DocumentedLive at oraclefm.com, built on Next.js and Netlify.
- DocumentedThe tools are marked alpha and beta on the site. An "AI Compliance Portal" page exists as a product concept with demonstration data; it is not a working portal.
What is documented
- UnknownNo outcome, metric or business result is documented for this work.
What this demonstrates
- InferencePositioning, content model, publishing workflow, interface and implementation designed as one system, and a deliberate line between what is deterministic and what is probabilistic.
Each line is labelled by the evidence behind it. Documented: checkable in code, a live product or a primary artefact. Stated: his own account. Attributed: a wider team's or client's. Inference: a reading of the evidence, not a fact. Unknown: not established. No outcome or metric is documented for any of this work. The same record is published for AI assistants at /llms-full.txt.