My Key Solutions

✅ Redesigned the onboarding process (1/3)

To lower the learning curve, I designed 1️⃣ a landing page with a clear punchline, 2️⃣ an interactive guide, and 3️⃣ example prompts to help users get started quickly.

✅ Introduced AI-powered smart search (2/3)

I introduced natural language search with AI-powered intent parsing. Users can review and modify the AI-parsed query, while the original structured search is accessible under “Advanced Search.”

✅ Introduced AI assistant for result interpretation (3/3)

The original result was a complex knowledge graph. Now, an AI assistant offers reliable, reference-backed interpretations and responds with context awareness when users select elements in the graph.

02.
Redefine
research workflows

Dancing with Complexity

To redefine our platforms into an all-in-one, AI-based platform, I had more workshops with our researchers and started by mapping the original workflow and identifying AI touchpoints.

What are the AI touchpoints in current workflows?

The yellow-highlighted steps mark where we decided to introduce AI.

Too complex, let's simplifie it.

Redefining the workflow with AI

From question to insight—faster, smarter, and more intuitive.

03.
Design
Human-AI Interaction

Design with AI, and its limits

Balancing the power of AI with human control, trust, and transparency.

AI's Pros and Cons & Our Design Solutions

✅ AI's strength

Understand natural language questions.

❌ AI's limits

May misinterpret ambiguous input.

🙌 Design solution:

Surface the AI-parsed query and let users refine it.

✅ AI's strength

Retrieve and summarize knowledge from graphs and literature

❌ AI's limits

Can generate hallucinated or outdated content

🙌 Design solution:

Support traceability with cited sources.

✅ AI's strength

Suggestive query design

❌ AI's limits

Lack of precision for expert use cases

🙌 Design solution:

Manual override: switch back to traditional search mode for advanced query crafting

Prompt-oriented Component Design

Context Steering
A UI module that lets users select graph nodes or summary segments to inject into the prompt, enabling fine-grained control of dialogue context.
Boosts relevance, reduces ambiguity, and reinforces user agency.

More High-fidelity prototypes

04.
Takeaways

Designing for AI-first Products

In this project, I approached the Large Language Model not just as a backend engine, but as a design material with unique properties: unpredictable, context-sensitive, and deeply conversational.
‍
‍
To support effective human-AI interaction, I designed a system where:
‍
💬 Context becomes the interface — Users shape what the AI knows through intuitive selection of graph nodes and summaries.
‍🤖 Prompting is embedded in interaction — Instead of typing prompts, users engage with structured components that generate meaningful inputs under the hood.
‍🩵 Trust is built through transparency and control — By letting users steer context and inspect what the AI sees, we restore agency in a probabilistic system.

‍
This is AI product design that goes beyond prompting — it’s about crafting semantic environments where AI and humans co-create meaning.

Blog

🤨

The key functions are scattered across 4 platforms. Researchers waste their time switching among them.

🧐

Hard to Read

“The results are all graphs and dense papers. I don’t know where to look.”

🤯

Not Smart enough

“I often struggle to tell the platform what I’m actually searching for.”

AI-driven Data platform

Project Brief

Client

Project Type

timeline

team

1. Overview‍

1.1 Challenges

1.2 Goals

1.3 Solutions

1.4 Impact

02. Design process

2.1 Balancing different users' needs

2.2 Build product that can scale

03. Key takeaways