Technical Demonstration · Portfolio Project
A working demonstration of what happens when you give a content team AI assistance they can actually trust: structured content, grounded answers with sources, and a human approving every publish.
9
AI assistant operations
Outlines, summaries, SEO, claim screening
7
Evaluation dimensions
Grounding, relevance, safety, a11y…
8
Seeded articles
Fully written, structured corpus
10
Typed API endpoints
Consistent envelopes, validated inputs
I built this as a technical demonstration for a Forward Deployed Engineer application. The architecture, workflow engine, APIs and tests are real implementations. Anything simulated (AI telemetry, the demo LLM) is labelled as such. No clients were involved.
From the platform
Governance is not a document you write after launch. It is a set of controls wired into the workflow itself: who may approve, what gets logged, and what a model is never allowed to do alone.
Bolting a vector database onto an app does not give you grounded answers. Retrieval-augmented generation works when the corpus, the chunking strategy and the contract between retriever and generator are designed together.
Most migration pain comes from moving content before modelling intent. Get the content model right and the move becomes mechanical. Get it wrong and you pay for a decade in validation scripts.
The challenge
Most organisations have plenty of content and no easy way to find any of it. Editors rewrite things that already exist. AI drafting shows up before anyone has decided who approves it or where it came from.
The solution
Content lives behind one typed interface. The AI helps draft and answer questions, every answer cites its sources, and nothing gets published without a named person approving it. The audit log records all of it.
How it works
An editor starts — or asks the assistant for — an outline. Any AI artefact is stamped AI GENERATED DRAFT from birth.
Summarise, adapt per audience, suggest tags, screen unsupported claims, pre-check accessibility.
Seven dimensions scored automatically. Results are advisory — they inform, never replace, the human decision.
A named editor reviews flagged spans, edits, then sends for review or rejects with a reason.
An approver from the allow-list signs the exact revision. The event lands in the audit log.
Only approved drafts can publish — and never by the AI orchestrator itself.
RAG-powered discovery
Ask a question and the system finds the relevant passages in its own index, hands them to the model as quoted data, and shows you every source it used. If coverage is thin it says so instead of guessing.
Open AI Discovery →AI governance
Accessibility
Semantic landmarks, ordered headings, keyboard-complete flows, 44px touch targets, visible focus rings, reduced-motion support and status that never relies on colour alone.
Performance
Every public page pre-renders at build; interactions are small client islands calling typed JSON APIs at the edge. No web fonts, no trackers, minimal client JavaScript.
Technical capabilities
ContentPort abstraction over Contentful or the local seed store — one interface, swappable backends.
Nine assistant operations behind a labelled-draft contract. Output is never auto-published.
BM25 retrieval, delimited untrusted context, grounded answers with inspectable citations.
Draft → evaluate → review → approve → publish, enforced in code with an append-only audit trail.
Seven-dimension scoring gating every AI draft before it reaches a human reviewer.
OpenAI-compatible protocol (DeepSeek, OpenAI, self-hosted) with circuit breaker and demo provider.
Structured health endpoint, per-operation token accounting, simulated spend dashboards.
Semantic HTML, 44px targets, visible focus, reduced-motion support, WCAG-conscious contrast.
Static export, hashed assets, edge worker APIs — fast on mid-range phones and modest networks.
Open the discovery console, run an AI assist through the editorial workflow, and inspect the governance dashboard — no keys, no setup.
Part of the Vuva Systems portfolio — vuvasystems.com