FitzStyles is a fashion-technology platform offering an AI-powered personal styling experience across web, iOS, and Android. The product serves both end consumers (B2C) and fashion brand partners (B2B), bridging the gap between what people own and what they actually wear — while giving brands a direct, measurable link between AI-driven styling recommendations and confirmed sales.

Business Challenge

Most People Own the Clothes. They Just Can't See the Outfit.

Decision fatigue is real. The average person owns enough clothing to construct hundreds of distinct looks — but without a trained eye, those combinations remain invisible.

Three compounding problems defined the challenge Prognos Labs was brought in to solve:

• Daily decision burden: Users struggle to create outfits from clothing they already own, leading to underutilisation of existing wardrobes.

• Product-grounding failure: Generic AI fashion tools hallucinate brands and products that do not exist, destroying user trust.

• Discovery gap: Fashion brands remain invisible at the precise moment a customer is most ready to buy, with no causal link from AI interaction to sale.

Traditional personal styling costs hundreds of dollars per session — making professional advice inaccessible to most consumers. Prognos Labs was engaged to close all three gaps simultaneously: reduce the daily decision burden, ground AI recommendations in real inventory, and build a commercially accountable discovery layer for brand partners.

Solution

One Platform. Three Surfaces. One Shared Intelligence Layer.

Prognos Labs conceived and built FitzStyles from scratch — a production-grade AI platform spanning three surfaces unified by a single data layer. The same preference store, embedding index, and attribution pipeline power all three.

Surface 1 — Fitz Chat: The AI Stylist

Fitz Chat is the primary interface — not a chatbot but an orchestrated system of specialist agents, a real-time embedding index, and a persistent memory layer working in concert. Every outfit recommended is retrieved from the live product catalog via vector similarity search and hard structured filters — never invented.

Surface 2 — The Social Platform

A full social layer where users post outfits, follow creators, and discover looks from people with similar body types and style profiles. Every post is shoppable: Fitz's recommendation engine maps tagged items to live catalogue inventory, surfacing affiliate or brand-direct purchase links automatically.

• Shoppable outfit posts — users tag garments to live SKUs; discontinued links auto-resolve to catalogue alternatives.

• Creator profiles with follower counts, style affinity scores, and a verified badge system for brand-partnered creators.

• Saves and collections — named boards where saved items feed back into Fitz as implicit preference signals.

Surface 3 — The Ecommerce Storefront

A native storefront embedded directly in the app — not a redirect to a third-party retailer. Users browse, filter, add to cart, and check out without leaving the platform. Every product is drawn from the same catalogue that powers Fitz recommendations, keeping inventory and pricing always in sync.

• Personalised storefront.

• Native cart and checkout with full purchase flow within the app.

• Size-aware filtering: out-of-stock sizes suppressed entirely.

• Brand storefronts — each partner has a dedicated shop-in-shop managed from their B2B dashboard.

• AI-written product descriptions adapting tone based on the user's stored profile.

• Wishlist-to-recommendation loop — wish-listed items inform future Fitz recommendations.

Want to build Custom AI Software? Let's talk

Architectural Decisions

Every technology decision was governed by a single constraint: ship software that works in production for real users, not a proof of concept that impresses in a slide deck.

Multi-Agent over Monolithic Prompt

Prognos Labs designed a directed graph of specialist nodes rather than a single large prompt. An intent classifier running at temperature zero routes each message deterministically. Each specialist node is cheaper, faster, and independently testable. The result is a system that becomes more accurate and more cost-efficient the more it is used.

One Database, All Layers

The preference store, embedding index, vector search, auth, and file storage all run on a single Supabase Postgres instance with Row-Level Security. No secondary vector database. Row-Level Security isolates all user data at the database layer — simpler to operate and cheaper to scale.

Process of Delivery

From Zero to Production in 11 Weeks

FitzStyles did not exist before Prognos Labs. The team conceived the product architecture, designed the full system, wrote every line of code, and shipped a production-grade AI platform spanning web, iOS, and Android — in under three months.

What Was Delivered

AI & Intelligence Layer

• Multi-agent orchestration with specialist nodes

• Advanced RAG pipeline

• Vision inference for outfit photo feedback

• Persistent preference memory— propagated as RAG filters on every turn

• Full observability — every agent step logged and post-hoc debuggable

Platform & Infrastructure

• Web application

• Native iOS and Android apps from the same codebase

• Supabase auth, Row-Level Security, storage, and Postgres

• 8 B2B brand dashboard pages: catalog, campaigns, analytics, messaging

• End-to-end purchase attribution pipeline (view → click → cart → purchase → refund)

• Production CI/CD deployment pipeline

Data Privacy & Security

• Row-Level Security on all Supabase Postgres tables ensures complete user data isolation at the database layer.

• Supabase Object Storage with scoped access policies for user-uploaded wardrobe images.

• Auth handled entirely via Supabase — no custom credential management.

• All preference and behavioural data are stored within the single Postgres instance, reducing third-party data exposure.

• LangSmith observability retains full trace history for audit and debugging without exposing user PII externally. 

07  Success Metrics

The figures below reflect the platform's performance at launch across all three surfaces.