The problem with one model
When we first built Smart Island, everything ran through a single Azure OpenAI deployment. Job classification, advisory reports, image generation, news analysis — all the same model, all the same endpoint.
It worked. But it was like hiring a brain surgeon to also do your filing.
Large language models have different strengths. Some are fast and precise at structured tasks. Some are brilliant at creative, opinionated analysis. Some generate stunning images. Using the most expensive, most capable model for every task is wasteful — and often produces worse results than a cheaper model that's better suited to the job.
Four models, four jobs
As of today, Smart Island's 65+ automated pipelines use four different AI models, each deployed through Azure AI Foundry:
GPT-5.4-mini — Job enrichment
Every job vacancy scraped from the Isle of Man goes through an AI enrichment pipeline: SOC2020 classification, skills extraction, automation risk scoring, O*NET mapping, salary estimation. This is high-volume, structured work — hundreds of jobs processed daily.
Why this model: GPT-5.4-mini is fast, cost-effective, and excellent at structured extraction. It follows classification taxonomies precisely, which is exactly what you need when mapping job titles to SOC codes. Using a larger model here would be 10x the cost with no quality improvement — the task is well-defined enough that a smaller model handles it perfectly.
Grok-4.3 — AI advisors
Our 9 weekly advisors (Careers, Economics, Employer, Strategy, Sector, Biosphere, Economy Stats, Subsea, and Vannin) need to do something fundamentally different from job classification. They need to have opinions.
The best advisory content doesn't just summarise data — it takes positions, challenges assumptions, and says what others won't. Our prompts explicitly demand this: "tell employers what they need to hear, not what they want to hear" and "if the government does nothing, X happens."
Why this model: Grok-4.3 is naturally more direct and willing to take positions. It doesn't hedge with "it could be argued that" — it argues. Combined with our "Signal → So What → Now What" prompt framework, it produces advisory content that reads like a sharp consultant's briefing rather than a committee report.
GPT-image-2 — AI illustrations
Smart Island has 200+ AI-generated images: bird illustrations, candidate portraits, conservation species, brand graphics, ecosystem visualisations. These are generated through dedicated image pipelines and served as static assets.
Why this model: GPT-image-2 produces significantly cleaner, more detailed illustrations than its predecessor. Our flat-design geometric avatars for election candidates went from "recognisable" to "genuinely good" with the upgrade. The style consistency is better, gradients are smoother, and the model follows complex prompts more faithfully.
GPT-5.6-sol — Fallback / complex reasoning
For tasks that need deep reasoning but not the opinionated edge of Grok — complex cross-referencing, Moody's credit analysis validation, and one-off analytical pipelines — we keep GPT-5.6-sol available as a precision instrument.
Why this model: Some tasks need careful, methodical analysis rather than strong opinions. Cross-referencing a credit agency's claims against actual government data requires precision, not personality.
How Azure AI Foundry enables this
The key architectural decision was moving from a single Azure OpenAI endpoint to Azure AI Foundry's model catalogue. Foundry gives us:
Model marketplace access. GPT, Grok, Claude, DeepSeek, Mistral, Cohere — all deployable from the same console, all billed through the same Azure subscription. We can test a new model in minutes, not days.
Serverless API deployments. Marketplace models like Grok deploy as serverless endpoints — no GPU reservation, no idle costs. You pay per token, and the endpoint scales automatically. For weekly advisor runs that take 20 minutes and then sit idle for 7 days, this is dramatically more efficient than a reserved deployment.
Unified authentication. Every model, regardless of provider, authenticates through Azure's identity layer. One set of API keys, one set of network policies, one audit trail.
Side-by-side testing. When we evaluated Grok vs GPT-5.6 for advisors, we could run both against the same prompts with the same data injection and compare outputs directly. The Grok outputs were measurably more opinionated and engaging — exactly what our advisory content needed.
The architecture
Our scraper package maintains a shared AI client module (azure-ai.ts) that abstracts the model routing:
getAzureClient()— returns the enrichment client (GPT-5.4-mini via Azure OpenAI)getAdvisorClient()— returns the advisor client (Grok-4.3 via Foundry serverless)- Environment-driven switching — change a model by updating env vars, no code changes needed
The advisor client automatically detects whether it's connecting to an Azure OpenAI deployment or a Foundry serverless endpoint, so the downstream pipeline code doesn't need to know or care which model it's talking to.
What changed in the advisors
Switching models was only half the upgrade. We also:
Built a news digest pipeline. Previously, advisors only received raw headlines — 25 bullet points with no context. Now, a dedicated news-digest pipeline scrapes 7 IoM news sources plus Google News, fetches article body text, and uses AI to produce a structured weekly digest with top stories, emerging themes, political context, and community pulse. Advisors get understanding, not just headlines.
Overhauled Vannin. Our master advisor, Vannin, has been reimagined as a maverick backbench MHK — the kind of politician journalists love to quote because they say what everyone is thinking but nobody in government dares to say. Each of Vannin's 10 perspectives (from Optimistic to Manx Crab) now reads like a speech in the Chamber, not a committee report.
Raised the temperature. Literally — advisor generation now runs at higher temperature settings (0.5–0.75 vs the previous 0.35–0.4), letting the model take more creative risks in its analysis and commentary.
The lesson
Don't treat AI models like commodities. They have different personalities, strengths, and cost profiles. The right architecture lets you match each model to the task it's best at — fast and precise for classification, opinionated for advisory, creative for images.
Azure AI Foundry's model catalogue makes this practical. Deploy four models in an afternoon, route traffic based on task type, and swap models by changing an environment variable. No code changes, no infrastructure migration.
For a platform that runs 65+ automated pipelines across 35+ datasets, that flexibility isn't a luxury — it's how you build something that actually works.
Smart Island is the Isle of Man's open data platform — 35+ datasets, 93+ MCP tools, and 140+ interactive pages. Built by Manx Technology Group.