ML/STUDIOSresearch-grade applied AI/ML
All work
Case study · AI assistant · Fine-tuned model

Raven AI Stylist

A conversational shopping stylist that gives personalised outfit advice and shows how a look might appear with AI-generated try-on images — powered by a model trained specifically for fashion.

Live Multi-agent Virtual try-on

Overview

Raven shows that a model trained for a specific domain, wired into a well-designed multi-agent system, can make recommendations feel genuinely personal.

You chat with a stylist that understands your taste and occasion, get concrete outfit suggestions, and see AI-rendered try-on images of the looks — not just descriptions. Behind the scenes a general model runs the conversation while a fashion-specialist model handles the expert advice, so replies stay both natural and on-point.

How it works

  1. 1

    A model fine-tuned for fashion

    A base open model (GPT-OSS-20B) is fine-tuned with LoRA (via Unsloth, 4-bit) into a domain adapter — "sigmoi" — that gives sharper, more relevant style guidance than the general model alone.

  2. 2

    Orchestrator + specialist agents

    A stateful orchestrator on the general model coordinates the conversation and calls stateless specialist agents (built with LangGraph): the base model routes and the sigmoi adapter specialises — same endpoint, selected by name.

  3. 3

    AI virtual try-on

    A dedicated agent composes a prompt and renders a photorealistic try-on image with Gemini Flash Image, so you can see the outfit rather than imagine it.

  4. 4

    Served in the cloud, streaming

    The model runs on an OpenAI-compatible vLLM endpoint on AWS; the app streams replies token-by-token for a real-chat feel.

  5. 5

    Graceful degradation

    If the inference endpoint is unavailable, the app stays up and the stylist shows a warm "temporarily unavailable" message — the rest keeps working.

Architecture

An orchestrator + stateless-sub-agent pattern. The browser talks only to the orchestrator; the orchestrator calls the sub-agents privately.

Browser ─ CloudFront (raven.mlstudios.ai) ─┬─ / → S3 static site (Next.js) ├─ /api/* → Orchestrator (streaming) └─ /users* → Profile API │ Orchestrator (base model) ─ invokes ─┬─ Style Agent (sigmoi) └─ VTO Agent (sigmoi) ─ Gemini (image) │ └──────── sigmoi-api (vLLM, OpenAI-compatible)

Built with

Python 3.12GPT-OSS-20B LoRA / UnslothLangGraph vLLMAWS Lambda (arm64) CloudFront / S3DynamoDB Next.jsGemini Flash Image Terraform