Trade Fair Data Platform

The Challenge

For an exhibition stand construction company, the calendar drives the business: setup deadlines, deadlines for technical guidelines, opening hours, teardown dates. This data comes from hundreds of organizers and exists in completely different forms — running text, tables, changing layouts, multiple languages, sometimes buried in downstream documents.

Until now, everyone researched it themselves. Whenever a project came up, the responsible project manager collected the dates — often for the same trade fair as colleagues doing exactly the same thing for their own project. For popular trade fairs, this meant up to five project managers working on the same research, each spending 15 to 30 minutes, across what is now more than 650 relevant trade fairs. And those were just the dates. A trade fair’s technical guidelines regularly run to 10 to 15 pages, read like terms and conditions, and are densely packed with requirements — analyzing them costs a multiple of the schedule research.

More expensive than the time spent was the inconsistency: results differed from one another. Different interpretations, different versions, different sources. In the end, two project managers would be debating which setup date was actually correct.

And it did not stop at debates. Deadlines were missed. Materials and transport were booked based on incorrectly researched dates — with all the follow-up costs that triggers in a deadline-critical trade. Ongoing maintenance made it worse: dates shift, deadlines get moved, new event years are added. A spreadsheet researched once quietly goes stale.

On top of that comes a second group of users with an entirely different interest: sales. For them, a trade fair is not a deadline but an opportunity — knowing that a fair takes place in a few months means being able to approach companies planning a stand there in good time. The gap here was less the approach than the foundation: a complete, current overview of the relevant trade fairs and a reliable prompt at the right moment. Both previously lived in individual lists and in the experience of a few people.

What was needed, therefore, was not just automation but a single, reliable source of truth that both groups of users could rely on.

Our Approach

We built a fully serverless platform on AWS — exclusively for this customer, who deliberately wanted the solution as their own competitive advantage rather than as a standard product on the market. At its core is a multi-stage processing pipeline in which two AI agents with different strategies extract the relevant information from the source material, transform it into a unified data schema, and continuously check it for deviations.

AI agents: two strategies instead of one

Classic rules or templates fail on source material whose structure, wording, and level of detail differ completely from case to case. We therefore rely on LLM-based extraction against a fixed field schema — and deliberately built two agents with different approaches that back each other up:

• Primary agent — broad capture in a single pass: Covers the available source material in breadth and extracts the target data from it. Fast and cost-efficient for the standard case, operating under a hard budget cap per run.
• Fallback agent — iterative, targeted search: An in-house development in Python. Instead of processing everything at once, it checks after each round which mandatory fields are still missing and decides specifically where to continue searching. It runs without a credit limit and only stops once the data is complete or the maximum number of iterations is reached — precisely the cases where the primary agent exhausts its budget.

If the primary agent fails, the fallback agent automatically takes over the same job. The chain is deliberately limited to two links and cannot run into endless loops.

Operations show that the effort for the second agent pays off: before it was introduced, a consistent 20 % of trade fairs remained unprocessed — they simply hit the primary agent’s budget cap and had to be reworked manually. These are exactly the cases the iterative agent now handles.

Further characteristics of the AI layer:

• Interchangeable models: Model access through an abstraction layer, model choice configurable per run — no lock-in to a single provider
• Prompts and field schema as configuration: Domain adjustments to prompts, extraction fields, and instructions without rebuilding the agent library
• Traceability: Every run returns not only the result but also where the information was found, the number of iterations, the reason it stopped, and the cost incurred
• Container-based deployment: The fallback agent runs as a Docker container Lambda from ECR, because its processing environment does not fit into a classic Lambda package

Why the data can be trusted

An AI that delivers dates is only worth as much as the trust placed in its results. The platform is therefore built so that automatically captured data remains verifiable and correctable:

• No silent overwrites: If an automated run finds a value different from one a user previously set manually, the deviation is applied and reported as a conflict. Users are alerted precisely to those cases where their correction and the new finding diverge — instead of being quietly overruled or, conversely, left sitting on an outdated value.
• Gaps are made explicit: Every trade fair carries a completeness status. Incomplete records are visible and can be addressed specifically — nobody works with partial data unknowingly.
• Every entry is traceable: For each result, the system records where the information came from and how it was produced.
• One version for everyone: Instead of parallel individual research, all roles access the same dataset. Debates about conflicting dates disappear.

Architecture highlights

• Asynchronous processing pipeline: Decoupled Lambda stages via SQS with job tracking, backoff, retries, dead letter queues, and automatic handover to the fallback agent when the primary path fails
• Structuring unstructured data: Conversion of heterogeneous source formats — including PDF — and extraction of the business-relevant dates and deadlines into a normalized schema, including multiple time ranges per field
• Data quality model: Separate dimensions for technical processing status and business data completeness (INCOMPLETE / MINIMUM / COMPLETE), so users immediately see which trade fair still needs manual work
• DynamoDB single-table design: All entities in one table, access patterns via global secondary indexes — tenant isolation, subscriptions, due dates, and cascading deletes without scans
• Change detection: New results are compared against the existing record; shifted dates and conflicts with manual corrections trigger targeted notifications
• Infrastructure-as-code with AWS CDK: Multi-stage environments plus short-lived, fully isolated development environments at the push of a pipeline button
• Testability despite AI: An interchangeable mock provider delivers deterministic results through the same interface as the real agents. This makes it possible to lock down every use case — including the error scenarios that rarely occur in live operation — as repeatable, automated tests, securing product quality over the long term

Two groups of users, one data foundation

Project managers and sales look at the same trade fair from opposite directions. Instead of building two tools, every user gets the same dataset — but their own lead time:

• Project managers plan backwards from the date. They are reminded of setup deadlines and the expiry of technical guidelines, typically two months in advance.
• Sales plan forwards towards the opportunity. They are notified months before the fair begins — the lead time is freely configurable — and can approach companies planning a stand there with sufficient lead time and in a coordinated way.

What matters is that every user subscribes to their own trade fairs and configures the reminders themselves — sales and project management alike. A shared data foundation thus becomes a personal work queue for each individual, without anyone centrally distributing lists or tracking who needs to be reminded of what and when.

Key features

• An elegant interface that works without training: A modern, uncluttered interface in which status values, roles, and actions are consistently visualized as icons and comprehensively backed by tooltips. Users discover the functionality while working rather than looking it up in a manual — for an application opened only a few times per week, this determines adoption. The interface receives consistently strong feedback from its users
• Reminders instead of checking: Subscriptions per trade fair and role with freely configurable lead times, multiple reminders per fair possible
• Role-appropriate views: Permission model with four roles, propagated via Cognito and JWT
• Self-service single sign-on: Administrators connect their Microsoft Entra tenant via OIDC themselves — without redeployment; existing accounts are linked automatically on first SSO login
• Automated season rollover: Expired trade fairs are archived and automatically re-captured for the next event year
• Event-driven notifications through a dedicated queue, decoupled from the request path
• We don’t stop at deployment: Monitoring and alerting are an integral part of the solution — CloudWatch dashboards and metrics for every processing stage, alarms forwarded to Microsoft Teams. Incidents reach us before users report them

Results

• Over 650 trade fairs are kept up to date automatically — without anyone maintaining them by hand
• Duplicate research eliminated: Depending on the project situation, the same trade fair was previously researched in parallel by up to five project managers, at 15 to 30 minutes per pass. Today everyone works from a shared version — deviations are reported as conflicts instead of being adopted unnoticed
• Missed deadlines and incorrect orders eliminated: Date changes are detected and actively reported before materials and transport are booked against them
• Sales can engage early and in a coordinated way: a complete overview of the relevant trade fairs plus a timely reminder per fair — subscribed and configured by each sales representative individually
• Around 20 % of previously unprocessable trade fairs are now covered — the iterative fallback agent closes exactly the gap left by fixed processing limits
• In productive use since 2025 and continuously extended — the customer keeps commissioning new features
• Implemented exclusively for the customer — the platform remains their competitive advantage and is not passed on to third parties

Outlook

The next expansion stage tackles the most labor-intensive remaining task: analyzing the technical guidelines.

These documents are the exact opposite of a data sheet. 10 to 15 densely packed pages per trade fair, written in the style of terms and conditions, covering requirements for stand construction, materials, safety, logistics, and deadlines. Only part of it is relevant to the customer at all — and that part all the more so. What makes it harder is that individual requirements change from one event year to the next: anyone who merely skims the document because it was “already read last year” misses precisely the change that matters this year.

The analysis is correspondingly demanding today — considerably more time-consuming than the 15 to 30 minutes of schedule research, and that across more than 650 trade fairs. This is exactly where the next AI-powered analysis step comes in: surfacing the requirements relevant to the customer instead of reading every document in full, and making deviations from the previous year visible instead of letting them slip through. Of all the expansion stages, this is the one with the greatest leverage.