Onliest REX
Designing REX — a luxury AI stylist and product-design engine for Onliest that turns vague fashion instincts into personalized, purchase-ready saree & blouse recommendations.
Figma · Notion · Claude AI · Jira
Product Owner (me) · CTO · AI/ML Lead · GenAI/RAG Lead · Fashion Curator · Security Lead
Business Requirements Document · Conversational & Wizard UX · Design System · Admin Governance Model
A Personal AI Stylist, Not a Search Bar
REX was built to make discovering and designing the perfect saree feel effortless, personal, and inspiring — for a platform where trust and craftsmanship matter as much as technology.
Onliest REX is the AI recommendation, personalization, and product-design engine for Onliest.ai — a luxury fashion platform. Rather than behaving like a generic e-commerce recommender, REX is built to act as a personal AI stylist: understanding a customer's body, occasion, taste, and budget well enough to design a saree or blouse with them, not just for them.
My Role
Owned the Business Requirements Document end to end and translated it into the conversational, wizard and try-on experiences shown here.
The Challenge
Turn a multi-agent AI system into something that feels like working with a stylist, not filling out a form.
The Approach
A low-friction conversation paired with a structured wizard, backed by a scored engine a customer can trust and an Admin team can control.
Fashion Discovery Was Guesswork
Luxury customers deserve a stylist's judgment, not a filter panel — but building an AI stylist that's trustworthy, explainable, and commercially sustainable is a much harder problem.
Premium fashion customers often need guidance beyond simple filtering. They may not know what design suits their body, skin tone, occasion, or budget — and existing platforms respond with endless catalogs instead of understanding. Onliest needed an AI system that could ask the right questions, reason over trust-sensitive data, and still feel like luxury, not automation.
Decision Paralysis
Browsing thousands of sarees with no reliable way to know what suits your body, skin tone, occasion or cultural context. Customers default to endless scrolling.
One-Size-Fits-All Recommenders
Generic e-commerce logic treats a saree, a blouse and jewellery the same way, ignoring that fit, drape and occasion rules differ per product line.
Inspiration Without Translation
Customers save Pinterest boards and wedding photos, but have no way to turn “I like this” into a manufacturable design without a human stylist.
Trust & Hallucination Risk
An AI stylist that invents availability, price or delivery destroys trust instantly — and bridal timing is non-negotiable.
An Agentic Engine Behind a Human Experience
Every conversational or Wizard interaction is backed by a governed, multi-agent architecture — designed so customers feel a stylist's intuition while the business retains full control.
REX addresses this with a tiered, agentic architecture: a conversational and guided-wizard front end backed by independent per-product-line scoring, tiered LLM routing, and hallucination guardrails — all configurable from a single Admin layer instead of buried in code.
Conversational + Guided Wizard
Free chat, voice or an uploaded image, or a structured 14-step wizard — two entry points into the same engine.
Product-Line Independent Scoring
Each category carries its own weights, fit rules and penalties, so a blouse's neckline scoring never distorts a saree's drape.
Hierarchical Agent Orchestration
An orchestrator routes work to specialist agents through MCP-style tools. No agent reaches the customer directly.
Tiered, Cost-Aware LLM Routing
Guests get rule-based results, unpaid users route to Llama, paid members to OpenAI — configurable by Admin, not hardcoded.
Hallucination Guardrails
Every product, price, inventory and delivery claim is validated against the catalogue before display.
Four Ways In, One Recommendation Engine
Whether a customer converses, steps through a Wizard, uploads inspiration, or simply browses, every path resolves to the same scored, explainable recommendation.
Interactive Stylist Chat
A live conversation with REX — via text or voice — where customers describe an occasion ('a saree for a birthday party in Miami around 8pm') and REX responds with a curated design plus a live Match Score.
Guided Design Wizard
A 14-step guided flow (fabric → border → pallu → blouse → fasteners) for customers who'd rather build their look step-by-step, with progress tracked and a live preview updating alongside every choice.
Upload Your Inspiration
Customers upload up to five inspiration images — Pinterest boards, wedding photos, celebrity looks — and REX extracts color, motif, fabric, and embroidery signals to recommend a similar, Onliest-suitable (not copied) design.
Live Virtual Try-On
Recommendations render on a mannequin scaled to the customer's real proportions, with the Match Score panel breaking down Confidence, Color Harmony, Occasion Match, and Preference Match in one glance.
Explore Curated Collections
For customers who want to browse rather than converse, curated designer collections are ranked by REX against the same fit, palette, and occasion profile — with full filtering by material, occasion, price, and color.
Design Studio Handoff
Any design — from chat, Wizard, or Explore — can be sent to Design Studio, where borders, drape, palette, and motifs remain fully editable until it's no longer REX's suggestion, but the customer's own design.
The working drawings
How a request actually moves through the system, and the problems the BRD was written against.
- 01Entry
- 02Orchestrator
- 03Discovery
- 04Scoring
- 05Guardrail
- 06Display
Search cannot hold intent
- A vague instinct has no keyword to type
- Generic search returns a catalogue, not a recommendation
- Customers abandon rather than refine
Unexplained results are not trusted
- A ranked list with no reason reads as arbitrary
- High-value purchases need a stated rationale
- Confidence has to be visible, not implied
Invented answers destroy trust
- Availability, price and delivery must be real
- Bridal timing is non-negotiable
- Every claim needs a source to check against
Orchestrated Agents, One Trusted Voice
No single agent talks to the customer directly. Every response is assembled by an Orchestrator and validated before it ever reaches the chat, Wizard, or Try-On panel.
Inside the REX Experience
Five core surfaces of the REX experience — from open conversation to structured design to discovery.
Interactive Stylist Chat
Conversational ModeVoice input, live transcript, and a stylist avatar sit alongside the generated design and a real-time Match Score — so the conversation and the recommendation are never more than a glance apart.
Guided Design Wizard
Structured Mode · Step 14 of 14A persistent step tracker and progress bar give customers who prefer control a transparent, undo-able path — while the same Match Score panel keeps both modes consistent.
Upload Your Inspiration
Image UnderstandingDrag-and-drop upload with format guidance (JPG/PNG/SVG/TIFF, 25MB max) sits next to REX's own conversational suggestion — reinforcing that upload is one of several ways in, not the only one.
Explore Curated Collections
Discovery ModeFilterable, ranked collections give browsing customers the same scored personalization as the chat and Wizard flows, closing the loop between passive discovery and active design.
REX Landing & Onboarding
Entry PointFour clear entry paths — Design Saree, Design Blouse, Just Exploring, Upload Inspiration — replace a single generic search bar, reducing the 'blank page' problem of a first-time AI stylist session.
Designed to Be Governed, Not Guessed
As a BRD-driven, pre-launch product, success here is defined by the constraints and metrics built into the design — the guardrails that make the AI stylist trustworthy at scale.
Course correction
What I got wrong
- First take
- I designed the conversational flow first and treated the guided wizard as a fallback for people who did not get on with chat.
- What changed it
- Walking the flows with the fashion curator made it obvious that a blank conversation is its own blank page. For a high-value purchase, plenty of customers want to be led — the wizard was not a fallback, it was a first-class path.
- What I did
- Gave both modes the same Match Score panel and the same underlying engine, so neither is the degraded option, and put four explicit entry paths on the landing instead of one prompt.
What This Project Taught Me
Designing an AI stylist for a luxury brand meant every UX decision doubled as a trust decision.
Designing for trust mattered more than designing for cleverness — customers forgive a clarifying question far more readily than a confidently wrong recommendation.
Giving every recommendation a visible, broken-down Match Score turned an opaque AI decision into something a customer (and the business) could actually reason about.
Two entry points — free conversation and a structured Wizard — served very different customer mindsets without forking the underlying recommendation logic.
Writing the BRD alongside the UX meant every screen was traceable back to a business rule, which made trade-offs (like tiered LLM routing) a design decision, not just an engineering one.
Want to work together?
Let's design AI experiences people can actually trust.