Tool selection
Does the agent map the user’s intent to the appropriate tool?
51WORLD
From a venue request
to a reliable action.
How can an AI agent turn a venue team’s request into the right action?
Venue operations and sales teams need to work with schedules, booking information and physical spaces. AIC brings an AI agent and interactive 3D tools together in an on-premises event-planning platform.
At 51WORLD, I translated user needs into product requirements and interactions, refined training data to improve agent reliability, and supported client validation and delivery.
- Role
- AI Product Manager
- Skills
- Requirements definition · Interaction design · Agent optimization · Client communication
- Time
- 2025–2026
- Tools
- Prototypes · Training datasets · Recurring test sets
My contribution Requirements and design proposals, data iteration, and client-facing documentation and demos, working with development teams and clients.
Workflow diagrams are reconstructed for this portfolio.
Make a business request specific enough to build.
I translated user needs into 10+ core intents for the event-planning agent. For the Shenzhen International Exchange Center, I designed calendar, booking-query and venue overview/detail interactions.
Review schedules
Venue calendar interactions.
Query bookings
Interactions for retrieving booking information.
Explore the venue
Overview and detail screens for spatial information.
Turn the proposal into a shared reference.
I used prototypes, documentation and working sessions to communicate proposals to developers, making the intended interactions concrete enough to discuss and implement.
The agent needed to choose and use the right tool.
Incorrect tool selection and parameter errors could prevent a request from becoming the intended action. I developed and refined training data for the locally deployed model to address both.
60%85%+
Combined intent-recognition and tool-call accuracy
I helped improve this measure through dataset optimization, evaluated against a recurring test set.
Parameter correctness
Does it supply the inputs needed to carry out that action correctly?
Build examples
Combine manually written examples with screened, model-generated data.
Check behavior
Evaluate tool-call correctness against the recurring test set.
Refine the data
Target incorrect tool choices and parameter errors in the next iteration.
Refined examples feed back into the next round of evaluation.
Help the client understand and use the product.
At the China National Convention Center, I supported on-site validation and delivery, and demonstrated AIC to the client’s sales team.
Client walkthroughs
Demonstrated the product and supported validation in the client setting.
User manuals
Documented product use so the client had a reference beyond the demonstration.
Chinese & English demos
Produced product demonstrations in both languages to explain the workflows.
My work connected feature definition, technical iteration and client enablement—bringing the product from specified interactions to supported use.