Fill in your details below and one of our AI experts will be in contact to find out how we can support you.
The Funny Finder tool that Datacom developed for the Melbourne International Comedy Festival is back for 2026 – once again helping audiences cut through hundreds of shows to find the acts that match their preferences. To unpack how Funny Finder was developed and what goes on under the hood, we sat down with Datacom Solution Advisor Clint Francis and Practice Lead Michael Waters, who shared what it took to build a safe, reliable, slightly irreverent AI-powered festival companion.
The brief for the Melbourne International Comedy Festival’s Funny Finder tool started with a deceptively simple goal: help festival‑goers navigate nearly 700 shows at the 2025 festival, then recommend the right gig based on what makes them laugh. Now in its 40th year, the festival is offering nearly 800 shows and the tool is helping people pick from the plethora of options.
Under the hood, it is a showcase for how to build a production‑grade, guard-railed, generative‑AI app on Microsoft’s stack without letting a large language model turn into an amateur comedian – or a compliance nightmare.
It made sense to anchor Funny Finder in Microsoft’s ecosystem, using Foundry, the Azure platform-as-a-service (PaaS) for building, managing and scaling enterprise-grade generative AI applications and agents.
“We employed the GPT‑family models as the core language engine. Rather than rely on a model’s baked‑in knowledge of comedy, the team deliberately used it as a linguistic engine, parsing natural‑language queries and generating fluent responses, but not allowing it to invent festival facts.”
“On the infrastructure side, we hosted Funny Finder on our private cloud infrastructure in its own dedicated Azure tenancy. Azure’s scalable architecture and Datacom’s rate‑limiting controls ensured the system could handle surges in traffic without either falling over or throttling genuine users into frustration.”
With 11th‑hour venue changes and show updates common, the development team ruled out heavy fine‑tuning. A static fine‑tuned model would rapidly diverge from reality as the programme shifted, and retraining cycles would lag behind the festival.
Instead, the team implemented a pattern similar to retrieval‑augmented generation (RAG):
Crucially, Datacom kept control of retrieval logic. For example, “more like this” recommendations and venue‑based suggestions ran through explicit filters and associations in the backend, not free‑form exploration by the model.
This prevented model‑hallucinated shows or mismatched acts stitched into plausible‑sounding, but non‑existent, line‑ups.
The team says one of the most important design choices was to “treat the model as untrusted until proven otherwise”.
“We wrapped the model with an internal agentic framework hosted on Azure Functions, which orchestrated retrieval, prompting and response validation. After the LLM generated an answer, validators checked the response back against the source dataset before it reached the user. When early iterations saw the model making up shows, the team tightened those loops until only responses consistent with the vector store were allowed through.”
Testing looked very different from traditional UAT (user acceptance testing). Because the same prompt can yield slightly different outputs, edge cases were harder to pin down and required aggressive “red‑teaming” by Datacom staff and the festival’s own team.
“Everyone threw their strangest prompts at Funny Finder, manually probing where guardrails broke or recommendation quality dropped, before taking advantage of Foundry’s automated evaluation tools.”
The Festival’s programme is intentionally edgy, which immediately put Funny Finder on a collision course with Azure’s built‑in safety filters. Summaries of adult or provocative show blurbs sometimes tripped content policies.
“We responded with a two‑pronged guardrail strategy. We kept Azure’s content filters in place for user prompts and outputs, so the chatbot would never generate responses the festival would regard as inappropriate. We also sanitised or pre‑processed some of the official show descriptions before they hit the model, to avoid triggering blocking rules while preserving the intent of the material.”
The development team also made a deliberate product call: Funny Finder should not be a comedian in its own right.
“During testing, when the model started riffing and becoming its own sort of comedian, we turned down the temperature and re‑aligned the prompts so the system behaved like a concierge rather than a stand‑up act. That single decision eliminated a whole class of risk where the model might freestyle its way into offensive territory.”
“To combat prompt injection attacks, we layered security around the application programming interfaces (APIs), web application firewall (WAF), API gateways, rate limits and monitoring to guard against both prompt‑based attacks and simple distributed denial of service (DDoS) attempts via the chat interface. The result was a bot that stayed firmly inside its designated lane, even as it handled open‑ended natural‑language requests.”
On the front end, Datacom’s web team built a clean, mobile‑first chat interface optimised for iOS and Android and embedded via the festival’s website. The user interface design stayed intentionally minimal to keep users focused on describing their tastes and exploring recommendations, rather than navigating a dense set of options.
Foundry’s telemetry, combined with Datacom’s own logging, captured rich interaction data – what people searched for, which acts attracted the most interest, and which prompts led to abandoned sessions. The festival tracked artists and shows of interest, while Datacom used the same signals to prioritise iterative improvements.
The whole solution was templated so it could be redeployed for other festival events. The team is now re‑engineering parts of the system to handle more diverse content while maintaining the same tight control over what the AI can say.
Funny Finder has become a repeat engagement for the Melbourne International Comedy Festival, delivered strong ROI on a lean budget. It has picked up awards – winning the top “Emerging Tech” award at the 2025 ARN Innovation Awards – and been featured in major sessions at events like Perth’s REMIX Summit.
More importantly for Datacom, it’s a living reference design for how to combine Microsoft Foundry, modern retrieval patterns and layered guardrails to ship AI that can play in edgy creative domains – and still keep it safe.
