TriloDocs

Designing structured AI workflows for Clinical Study Reports

TriloDocs is a platform that uses AI to generate Clinical Study Reports (CSRs) from uploaded clinical data tables and structured wizard inputs.
This project expanded existing AI capabilities by introducing a new platform features.

My role

Role

Product Designer

Scope and contribution

- Define end-to-end user flows
- Design all screens and interaction states
- Translate backend rules into UX logic
- Speficy validation behaviour
- Define version history and restore flows
- Ensure clarity between automation types
- Collaborate closely with engineering to align frontend and backend state models
project overview

TriloDocs is a platform that uses AI to generate Clinical Study Reports (CSRs) from uploaded clinical data tables and structured wizard inputs.This project expanded existing AI capabilities by introducing:

  • Structured table-based AI analysis (Advanced Search)
  • Guided AI-assisted drafting for narrative sections
  • A dual-path model separating deterministic suggestions from generative AI

The requirements were predefined. My role was to translate strict backend constraints into clear, state-safe user experiences.

The Design Challenge

The challenge

Translating complex backend constraints into a clear, predictable, and usable experience

The challenge wasn't defining what the system should do, it was how to design it:

  • Turning strict backend constraints into intuitive interactions
  • Designing branching logic without overwhelming users
  • Communicating regeneration rules clearly
  • Managing concurrent AI states
  • Preserving auditability and traceability
  • Differentiating deterministic logic from generative AI
Problem Definition, Roadmap & Milestones

Given the scope and technical complexity, the work was divided into three milestones.

Milestone 1: Advanced Search (Table-Based AI)

Enabled users to generate structured AI analysis from CSR tables with:

  • Comparative vs non-comparative flows
  • Multi-block input support
  • Strict regeneration rules
  • Concurrency handling

Focus: Input logic, regeneration governance, response lifecycle.

Milestone 2: Guided Narrative Generation

Extended AI assistance to free-text narrative sections with:

  • Structured configuration selectors
  • No free prompt injection
  • Regeneration tied to configuration changes
  • Full version history retention

Focus: Controlled flexibility and version transparency.

Milestone 3: Dual Path Model

Introduced a second pathway

  • Deterministic “Suggested Narrative”
  • Structured GenAI “Advanced Search”

Focus: Conceptual clarity and logic differentiation.

ideation

User Flows & Interaction Design

Milestone 1: Advanced Search

I mapped the full journey per table:
Select table → Configure analysis → Validate → Generate → Edit → Save → Regenerate (if allowed)

Key design explorations

  • Progressive disclosure to prevent cognitive overload
  • Dynamic rendering based on analysis type
  • Clear required vs optional input grouping
  • Explicit system states (in progress, blocked, unchanged input)
  • Messaging for regeneration constraints

Milestone 2: Guided Narrative

I mapped a new drafting journey:
Configure → Generate → Review → Compare versions → Edit → Restore → Save

Key design explorations

The main design focus was version management:

  • Clear tagging of generated vs user-edited versions
  • Restore workflows
  • Unsaved edit warnings
  • Immediate invalidation when configuration changes

The goal was making AI output reversible and traceable.

Milestone 3: Dual Path

Key design explorations

I explored how to visually and conceptually separate deterministic and generative logic

  • A clear analysis-type selection step
  • Distinct action buttons
  • Eligibility messaging per table type
  • Single-use enforcement states for suggestions
  • Consistent but differentiated output presentation
Final Design & Delivery

A guided narrative configuration panel with safe defaults

A dual-path interaction model separating deterministic and generative outputs

A dynamic branching input system for structured table analysis

A version history system with restore functionality

Clear regeneration and blocking states aligned with backend logic