Building the product that turns chaos into intelligence at the world's largest trade shows.
The events industry has a data problem nobody talks about. Before every trade show, organisers spend weeks chasing exhibitors for profile information, only to end up with inconsistent spreadsheets and copy-pasted website text. Visitors arrive to a search experience that amounts to scrolling an alphabetical list, or typing keywords that return whoever has the most text on their profile.
The real problem runs deeper: attendees don't know what to search for. They know the outcome they want — "a supplier who can handle our European logistics" — but can't translate that into a keyword. So they wander. Connections go unmade. The show floor underperforms.
A SaaS start-up came to us with a vision to fix this. They had the market relationships and a committed first customer event on the calendar. What they didn't have was an engineering team, a product, or a platform. They had three months.
Building rich, structured exhibitor profiles by hand is expensive, slow, and stale the moment it's collected. We needed accurate, detailed profiles generated automatically, at scale, with no human input.
At 9am on opening day, thousands ask questions at once, each triggering a chain of AI inference calls measured in seconds. Standard scaling breaks down, and the economics of an LLM per-request at that volume are ruinous.
A multi-tenant SaaS platform built from scratch, integrated with existing event-management systems, production-ready in three months — with no existing engineering team.
AI profiling pipeline
The system treats the open web as a data source: it crawls each exhibitor's public presence, condenses it, then runs parallel AI extraction across seven distinct dimensions of each business, each one resolved independently and normalised against the event's own taxonomy. Hours of manual research, produced automatically in minutes.
Semantic search for intent
Rather than matching strings, it embeds queries as vectors and searches semantically — understanding that "sustainable packaging supplier" and "eco-friendly materials manufacturer" are the same need. It deduplicates intelligently to surface the best exhibitors, returning ranked results with full context in real time.
Concurrency, solved architecturally
A semantic response cache — itself built on vector similarity — answers common questions from cache instead of fresh inference. As traffic spikes, the system gets more efficient, not less. Load-validated before every deployment, iterated to 100% reliability at peak.
We were the founding engineering partner — not handed a spec, but shaping the product alongside the founders. We owned architecture, infrastructure, the AI pipeline, integrations and performance — and established the engineering function itself: workflow, standards and hiring criteria that let the client build their own team with confidence.
The platform launched at the client's first event three months after kick-off. Over six months it was hardened and scaled to support five colocated events simultaneously — 65,000 users — at 100% API reliability under full load. The client had a production-grade SaaS product, an engineering team, and a platform built to grow with them.