01 Overview

Lost in the Graph

Guided navigation for complex, non-linear knowledge systems.

I designed a knowledge graph platform end-to-end. Its navigation remembers your path and intent.

  • Knowledge Graphs
  • Wayfinding
  • Intent-based UX
  • AI-assisted Search
  • Claude
  • Figma Make
  • Camera
  • Pen and paper
  • My brain

  • ~2×

    faster insight discovery

  • 40–50%

    fewer navigation steps

  • Less clutter

    in large graphs

  • CEO praise

    in executive review

The explorer: ask, explore, then retrace.

02 The problem

Too much graph, too little direction

  1. 01

    Structuring data

    Manual schema setup was slow and inconsistent.

  2. 02

    Non-linear navigation

    Where do I begin? How do I get back?

  3. 03

    Graph overload

    Large graphs became clutter and fatigue.

The problem wasn't visualisation. It was navigating a non-linear system.

03 Role & scope

End-to-end, from raw data to insight

  • Senior UX Designer

  • 2.5 months

    From December 2025

Decided together

  • Head of Engineering

    Rendering feasibility, scope, what waits

  • Product Manager

    Creation vs interaction, delivery order

  • Backend team

    Extraction, AI mapping, schemas

  • Frontend team

    Interaction, building the navigation

Two layers

  1. Raw data
  2. Creation layer
  3. Knowledge graph
  4. Interaction layer
  5. Insight

04 Creation layer · data → knowledge

One way in for any data

  1. Upload
  2. Extract
  3. Map
  4. Validate
  5. Generate graph

Structured · Upload

Start from a template

Explored

Connectors, as designed
  • Salesforce connectors and HRMS sync Next phaseDesigned and built, with scheduled sync. Parked for cost until the next phase.

Shipped

The upload step

CSV or .xlsx upload with a template. Files arrive in a known shape.

Decided with Head of Engineering · Product Manager

Structured · Map

Let AI draft the schema

Tried

The schema built by hand
  • Manual mapping, creating attributes on the spotSlow, and everyone built schemas differently.

Shipped

AI creates and maps; you confirm

AI maps columns into categories and attributes. People confirm or fix each one.

The trade-off: AI can be wrong. A person checks every mapping first.

Decided with Backend team

Unstructured · Extract

Context first, then extract

Tried

Everything extracted, regions boxed
Auto-drawn regions on an image
  1. Upload PNG / PDF / DOC
  2. Full extraction
  3. Mapping
  • Extracting everything in the fileSlow, costly, and mostly irrelevant.
  • Auto-drawn boxes you could redrawThey blocked the node dialog.

Shipped

Say what you're after first
  1. Upload PNG / PDF / DOC
  2. Insert context
  3. Extract as per context
  4. Mapping

First, people say what they need. AI extracts only that text.

Decided with Backend team

Unstructured · Map

Map with the context it came with

Tried

  1. Upload PNG / PDF / DOC
  2. Extract
  3. Map onto the structured data
  • Mapping onto the already-mapped structured data

Shipped

  1. Upload PNG / PDF / DOC
  2. Extract with context
  3. Map with that context
Structured and unstructured, side by side

Extract with context, then map. Better input, fewer tokens, lower cost.

Decided with Backend team

Structured files stayed simple. Each sheet shows what was mapped.

Progress you can read

05 Interaction layer · knowledge → insight

Ask, and the graph answers

  1. You ask
  2. It responds
  3. Nodes adapt
  4. You explore
  5. Edit if needed

Show me the Data Governance hierarchy.

Found it: Data Governance. Click the node to drill in.

Three systems keep it readable

  • Adaptive filters

    Built from your question

  • Relevance and clusters

    Only the relevant subgraph is drawn

  • Cross-linked

    Select a node; everything follows

Interaction · Explorer

From canvas and chat to explorer

Tried

Version 1
  • Version 1: a canvas beside a chatPeople lost track of where they were.

Shipped

You ask, the graph responds

History, breadcrumbs, filters and assistant work together. You always know where you are.

My call.

Explored: attributes spread round each category
Explored, not shipped: attributes folded into rings
Adaptive filters, live.
After: related nodes folded into clusters, with only the relevant links drawn.Before: a dense graph, every node and label at once.BeforeAfter
Drag to compare.

Explored, not shipped

  • Inline node editing

    Deferred for scope and time

  • Inline relationship editing

    AI relationships made it unneeded

  • Editing prompts and the AI's reasoning

    Inline, in the graph

  • Inline photos

    Dropped with photo extraction