All work

Case studyAWS Hackathon2026

Wear What You Own: A Multimodal Stylist

TL;DR

A fashion companion that turns closet chaos into curated confidence, using the items you already own.

  • Multi-agent architecture: 3 small AIs (Intake / Stylist / Behavioral), each with its own prompt + temperature + guardrails
  • Each agent is checkable and explainable on its own.
  • Morning decision time: 15 min → <30 sec · ~2.5× more outfit combinations from the wardrobe you already own
  • Solo hackathon build on multi-modal AI model (Amazon Nova Lite): end-to-end deployment → from user discovery to actual deployment

Open the app

the Stylemax prototype
Timeline
AWS Hackathon. Built in days, not weeks.
My Role
Problem framing, multi-agent architecture, prompt engineering, data schema, evaluation
Team
1 PM (me), solo build with AI-assisted development
Scope
3-agent system on Amazon Nova 2 Lite · tops, bottoms, outerwear and shoes · no accessories · Firestore backend

01

How I Decided the Path

What I picked, and what I said no to.

  1. 1

    Rejected: one model

    Its failures are opaque, it leaves no place for guardrails, and it cannot tell the user why.

  2. 2

    Rejected: shopping recommendations

    A tool meant to reduce buying pressure cannot quietly sell things on the side.

  3. 3

    Selected: three agents

    Each has its own prompt and temperature, guardrails sit at the seams, and every suggestion comes with its reason.

Intake: tag the closet

temp 0.2 · accurate categorization

Reads photos and user notes, and returns clean item tags: type, color, season and formality.

Stylist: compose three outfits

temp 0.7 · creative combinations

Returns item IDs only, never names. Unknown IDs are dropped before the user sees anything.

Behavioral: surface what's forgotten

temp 0.7 · neglected-item radar

Flags items unworn for 3 weeks or more. Wear history closes the loop: once an item is worn, its flag clears.

02

How I Validated

Kill criteria before the prompt. I wrote down what a failed outfit looked like before the stylist prompt existed, so the bar was not set by whatever the model happened to return.

  • Two bottoms is not an outfit. Discarded rather than shown. Every suggestion ships with a one-line reason, so a bad combination does not arrive quietly, it arrives argued for. Confident reasoning attached to a wrong answer is worse than no answer.

  • Two tops is not an outfit either, unless it is layered. An outer over an inner is a real choice a person makes. Two shirts is a bug wearing the same shape.

  • Invented items never reach the screen. The stylist returns item IDs only, never names, so anything hallucinated fails the lookup and is dropped before rendering. Outfits left with fewer than two valid items are discarded whole.

03

Cross-Agent Feedback Loop

The user's own behavior closes the loop: the AI gets out of the way the moment the unworn item gets worn.

Intaketags the closettemp 0.2Stylistcomposes 3 outfitstemp 0.7The userwears one, or taps Will tryBehavioralflags 3+ weeks unworntemp 0.7itemsreasonshistoryunworn items go back into the next outfits

More work

Sign-in and authorization for 60+ financial institutions, a palm-vein ID for 14 airports, and AI in research budgeting.

All work