My Role — Sole Designer & Product Lead

I owned this project end-to-end as a single-person design team, responsible for:
– Defining the problem and user needs
– Designing the full information architecture
– Mapping retrieval constraints into UX patterns
– Creating all interaction flows and UI states
– Building the design system
– Writing UX and implementation specifications
– Aligning RAG behavior with the UI and system logic

There were no additional designers, product managers, or UX teams involved — I handled the entire project independently from concept to delivery.

01 — The Problem

Smarteeva’s documentation had grown to thousands of MDX pages, spread across multiple versions, modules, and workflows.
Traditional keyword search wasn’t enough.

❌ What users struggled with

• Had to know the exact keyword

• Couldn’t find configuration steps buried deep inside docs

• Release notes lived separately and weren’t searchable

• Multi-version behavior confused new users

• Support teams answered the same repeat questions daily

Resulting business friction

• High support load

• Low adoption of documentation

• Slower onboarding for customers

• Internal teams constantly context-switching

The opportunity

Create an AI assistant that can:

✔ Understand natural-language queries
✔ Retrieve the exact relevant documentation
✔ Explain it in a readable way
✔ Provide citations
✔ Detect version mismatches
✔ Build trust at every step

02 — Users & Needs

Primary Users

• Implementation specialists

• QA engineers

• Support teams

• Internal developers

• Customer admins

Core Needs

• “Explain this feature simply.”

• “Where is this configured?”

• “What changed in the new version?”

• “What API parameter should I use?”

• “I need a trusted answer — not a guess.”

Smart Assist needed to be fast, trustworthy, readable, and deeply integrated into the workflow.

03 — Goals

Business Goals

• Reduce documentation search time by 80%

• Increase documentation adoption across teams

• Improve onboarding for new customers

• Reduce repetitive support queries

UX Goals

• High readability for long AI-generated answers

• Clear validation of trust (citations, version tags, confidence scores)

• Smooth navigation from answer → source document

• Scalable layout supporting technical depth

04 — Architecture Overview (Simplified)

Smart Assist runs on a hybrid RAG architecture tailored for documentation that changes constantly.

1. Stable documentation → OpenAI Vector Store

• SOPs

• API references

• Feature descriptions

Indexed once → fast, predictable retrieval.

2. Frequently changing docs → Pinecone Vector Store

• Release notes

• Version updates

• Changelogs

Dynamically updated.

3. Automated Update Pipeline

Triggered by admin:

1. Fetch MDX

2. Clean + chunk

3. Embed

4. Update Pinecone

5. Attach metadata (version, URL, type)

4. Unified Retrieval Layer

Parallel search → deduplication → Top-K merge.

5. LLM Layer

Structured answer generation with:

• Markdown

• Citations

• Confidence scoring

• Error awareness

• Version awareness

6. UI Delivery

• Scroll-optimized

• Readable formatting

• Clear fallback states

• Links to source pages

05 — UX Strategy

Key UX Priorities

1. Trust over creativity
   Citations, confidence scores, explicit version tags.

2. Clarity for long-form responses
   AI answers often exceed several paragraphs.

3. Predictable structure
   Location → Steps → Notes → Links.

4. Support multiple states

   • High-confidence answer

   • Low-confidence fallback

   • No results

   • Version mismatch

5. Minimal cognitive load
   Neutral surfaces + restrained color palette.

06 — Key Screens

Screen 1 — Onboarding & Quick Actions

A personalized welcome experience:

• “Hi Rhea 👋”

• Four primary use-cases surfaced as cards

• Gradient hero background to create hierarchy

• Clean dark UI for focus during long reading sessions

Screen 2 — High-Confidence Retrieval

User: “Where do I configure Complaint Intake Workflow?”

The assistant returns:

• Navigation path

• Step-by-step instructions

• Important version notes

• Source citations

• Reference links

Structured, predictable, and verifiable.

Screen 3 — Low-Confidence Retrieval

When context is insufficient:

• Yellow badge (“Low confidence: 0–20%”)

• Explanation of why

• Related searches

• Helpful links

• Red disclaimer that answers only reflect indexed documentation

This prevents hallucinations and builds long-term trust.

Screen 4 — Version Mismatch

User asks about v4.0, but feature exists only in v3.2.

The UI shows:

• Version difference badge

• “What to do in v4.0” alternative

• Mapping old behavior → new behavior

• Migration notes

This screen is unique — very few AI assistants handle version awareness.

07 — Design System

1. Design Principles

2. Color System

Optimized for dark mode & readability.

Primary

• Purple family (#6946FF, #7B5CFF) — actionable but calm

• Soft lavender (#EFEAFF) — focus rings & subtle surfaces

Neutrals
900 → 700 → 500 → 400 → 100
Inspired by Material elevation hierarchy.

Semantic Colors

• Green — successful retrieval

• Yellow — low confidence

• Red — retrieval failure

• Blue — info tags

3. Typography System

Inter

• High x-height

• Excellent readability

• Ideal for dense UI + long-form documentation

4. Accessibility

• 48px touch targets

• WCAG AA contrast

• Purple restricted to non-text areas

• Minimal motion to reduce fatigue

• Colorblind-safe palette validation

08 — Impact

Quantitative

• Search time reduced from minutes → seconds

• Documentation adoption increased across 4 internal teams

• Support escalations dropped

• Faster onboarding cycles

Qualitative Feedback

“This is the fastest way to understand Smarteeva.”
“Now I actually trust the answers — the citations changed everything.”
“The version mismatch detection is genius.”

09 — Challenges & Learnings

Working independently forced me to combine product strategy, UX thinking, system reasoning, and technical constraints into one cohesive solution.

Challenges

• Long technical answers required tight typography & spacing

• Multi-version behavior was difficult to surface clearly

• Some documentation had inconsistent formatting

• Needed to minimize hallucinations at all cost

What I Learned

• How to design retrieval-first interfaces

• How to blend AI system constraints with UX clarity

• How to structure conversations for technical accuracy

• How to build trust in enterprise AI tools

• How to collaborate deeply with engineering, data, and ML teams

10 — Flow Chart

11 — Final Reflection

This was a fully self-directed project — every UX decision, architectural insight, design detail, and deliverable was created solely by me.

Smart Assist became one of the most impactful internal tools in the product ecosystem.

It not only made documentation searchable — it made it understandable, verifiable, and trustworthy.

It strengthened my ability to design:

• AI-first workflows

• Complex enterprise UX

• Multi-version technical systems

• Scalable design systems