WRDBAn AI virtual wardrobe for iOS, and the data machine behind it.

We wear about 20% of what we own. The rest is invisible: forgotten in a drawer, lost behind the decision fatigue of every morning. WRDB closes that gap. You photograph a garment, the app removes the background and tags it, and from then on it builds outfits for you from what is actually in your closet, with a social feed where friends post their outfit of the day.

Three founders: Alex Matos is CTO, Pep Bustamante is CMO, and I am CEO and CFO. In practice I own data, machine learning, analytics and growth, and I wrote the first version of the app myself. Legal entity WRDB SCP, Barcelona. The app is live on TestFlight and the App Store review is ongoing.

The repository starts on 29 July 2025 with my base code. The first working outfit generation landed on 2 August, the database schema and cost-tuned agents on 1 August, and a Clean Architecture rewrite on 26 August: Presentation, UseCases, Domain and Data layers with dependencies pointing inward, so use cases test without a network or a simulator.

I am the original author of the DI container, the app entry point, the domain entities and validators, twelve repository protocols and their remote implementations. Since then my co-founder has carried most of the later refactoring, social features and pre-launch hardening while I moved to the data side.

The embeddings run on the phone. I built the PyTorch-to-Core ML conversion pipelines for FashionCLIP (a CLIP ViT-B/32 fashion encoder): download, wrap, convert to an fp16 mlpackage, validate parity against PyTorch, and generate the Swift integration. The model unloads after sixty seconds of inactivity to save memory.

The outfit generator scores color harmony, style consistency, occasion fit, weather appropriateness from WeatherKit across eight temperature bands, and vector similarity, evaluating up to 500 combinations per session with a diversity pass so the five suggestions are not the same outfit five times. A linear preference learner updates eight-dimensional weights from what the user accepts, rejects, modifies and wears, batched with a five-second debounce. Occasion learning and custom occasions shipped in April 2026.

In March 2026 I audited the analytics gap and shipped the fix in one day across fourteen phases: TelemetryDeck with 33 consent-gated signals (sessions, screen time, onboarding, outfit accepted or rejected or worn, ads, image upload, API latency, errors), a Supabase cron warehouse for DAU, retention cohorts, churn, bounce and the activation funnel, and a Next.js dashboard to read it all.

The dashboard is 32 pages on Next.js 14, Recharts and TanStack Query. One page is genuinely realtime: a live activity feed that subscribes over Supabase Realtime to signups, garments added, outfits posted, follows and affiliate clicks. The rest queries the warehouse on demand. Alerts do not live in the dashboard at all; they go to Slack through 29 edge functions, including a slash command that turns a natural-language question into a database query.

Marketing had to run without a marketing budget, so I built it as software. Six layers: a context brain of brand, audience and competitor documents; market intelligence; a strategy layer that writes a content calendar straight into Postgres; a content factory; publishing; and analytics that feed back into the brain. Total additional cost: 15 EUR of image credits, bought once, on subscriptions we already paid for.

Publishing is a cron every fifteen minutes that posts through the Meta Graph API and the TikTok Content Posting API with a UTM link per post. Daily pulls collect post and account metrics, detect anomalies (viral, underperformer, engagement drops, A/B winners) and score every post on a composite. Each post is tagged by hook style, emotion, call to action, visual style and trend, so the system learns which formula wins per platform. In June 2026 I added an auto-generator: a daily job that rotates a hook bank, renders reels from HTML with HyperFrames and Higgsfield, uploads them and queues them for an AI persona account. The first batch was twenty reels scheduled over two weeks.

The revenue model is affiliate commerce: recommend a product that fills a gap in someone's wardrobe, earn a commission when they buy it. I owned that side end to end, from the partnerships to the recommendation engine. In January 2026 I applied to the affiliate networks with a publisher profile and the pitch deck. CJ Affiliate blocked us because a promotional property needs a live App Store link; Awin España accepted the application on a refundable one-euro deposit. Inside Awin I applied programme by programme: thejewellershop.com ES joined on 29 January 2026, with a 16 percent conversion rate and 1.58 EUR earned per click on the network's own figures, then Bonheur Jewelry, Dooers and Zapatos OBI were accepted and Palomo Spain applied for.

A joined programme gives you a product feed and a tracked deep link per product. The jeweller's feed is a gzipped CSV of 14,274 products refreshed every four hours. A script downloads and parses it, a price filter from 2 to 150 EUR leaves 6,111, and a curation pass with category and brand exclusions, de-duplication on model number and a quality score selects the catalogue: 100 products at first, then 200 split evenly between women and men, and about 300 today across the programmes. Each product goes into the database with its affiliate link, its image with the background removed and its bounding box measured so it can sit in a flatlay next to the user's own clothes, tags from gpt-4o-mini for colour, style, material, season and occasion, a gender derived from the merchant category, and a 512-dimensional CLIP vector in a pgvector column, the same embedding the user's garments carry.

The marketplace has three tabs: For You, Missing in Wardrobe and Discover. One SQL function scores every product for the user in seven steps and returns a ranked page with a reason code. Two scores blend: a discovery score from global popularity, product quality, recency, click-through and a little exploration noise, and a personal score from CLIP similarity to the user's wardrobe, past interactions, wardrobe gaps by category, style, colour and price. How much of each depends on a richness score, a saturating function of how many garments, vectors, outfits and interactions the user has, so a new user sees discovery and a power user sees their own taste, with no threshold and no cold-start switch. Missing in Wardrobe raises the gap weight to 60 percent; Discover is the same feed for everyone. Dismissed products carry a penalty for thirty days, every page caps brands at three and categories at five, and every impression, view, click, save and dismissal is kept raw for training and rolled up by triggers and cron jobs into per-product and per-user statistics. Purchases arrive from Awin on a webhook that posts to Slack.