TotalLook

Inspiration

Online fashion still asks customers to imagine too much. A product may look great in a catalogue but feel completely different once it is combined with a person’s skin tone, proportions, existing wardrobe, or the rest of an outfit.

At the same time, most retailers cannot afford to build their own AI styling infrastructure, virtual try-on integrations, safety policies, and agent orchestration platform.

TotalLook was inspired by a different model: the customer remains inside the retailer’s experience, while TotalLook works behind the scenes as a business-to-business, agent-to-agent gateway. A retailer can connect its storefront or application to specialized styling agents without exposing the underlying complexity—or requiring the customer to understand AI agents at all.

What it does

TotalLook is a multi-tenant AI fashion infrastructure platform for retailers.

A customer begins inside a retailer-owned storefront and can provide a selfie, a full-body photograph, measurements, preferences, and the type of occasion they are shopping for. The retailer then sends a pseudonymous request to TotalLook through an A2A API.

TotalLook coordinates a complete styling workflow:

• Analyses visible skin-tone characteristics and creates a colour profile.
• Filters the retailer’s catalogue using gender, measurements, availability, category and inclusion policies.
• Ranks eligible garments using colour compatibility and customer context.
• Generates AI Clothes Virtual Try-On results.
• Supports footwear virtual try-on as part of the same experience.
• Returns the recommendation and generated preview to the retailer’s application.
• Escalates uncertain or sensitive recommendations to a human stylist.
• Records every important decision for auditing and traceability.

The platform supports both direct clothes and footwear try-on and a complete agent-driven styling journey. It also includes independent web and terminal clients to demonstrate that TotalLook is a reusable gateway rather than a single consumer application.

Retailers own the customer relationship, catalogue and frontend. TotalLook provides the intelligence, orchestration and integration layer behind them.

How we built it

TotalLook is built as a multi-tenant Ruby on Rails application using AgentKit for agent execution, workflows, memory, observability and human-in-the-loop review.

The system is divided into specialized agents:

• SkinDiagnosticAgent processes the customer’s image and produces colour evidence.
• InclusivityGuardAgent deterministically filters the tenant’s catalogue using real product metadata and customer constraints.
• GarmentSelectionAgent ranks eligible products and explains the recommendation.
• StylingVTOAgent routes garments and footwear to the appropriate YouCam virtual try-on capability.
• TotalLookCoordinatorAgent orchestrates the complete workflow and exposes it through an A2A contract.

We integrated YouCam APIs for skin analysis, AI Clothes Virtual Try-On and footwear try-on. The routing layer understands that clothes and shoes require different processing paths while presenting one consistent interface to retailer clients.

Each retailer receives an isolated account, catalogue, API key, usage limits and execution history. Requests use pseudonymous consumer references, allowing a storefront to access personalization without requiring TotalLook to know the customer’s real identity.

The project also includes:

• A retailer-style storefront demonstrating the real customer journey.
• An independent A2A web client.
• A terminal client showing protocol-level integration.
• A try-on playground for clothes and shoes.
• Human-review screens for low-confidence recommendations.
• Architecture, observability and execution-trace views.
• Deterministic fallbacks so an LLM failure does not break delivery.
• A synchronized cinematic demo that presents the experience, live agents and architecture together.

Challenges we ran into

One major challenge was connecting several asynchronous AI services into an experience that still feels immediate and coherent. Skin analysis, catalogue selection and virtual try-on have different inputs, response formats, processing times and failure modes.

Image handling was another important challenge. External services require publicly reachable images in supported formats. Localhost URLs, expired URLs and incompatible PNG files can cause a workflow to fail even when the agent logic is correct. We added validation, explicit failure reporting and catalogue health suggestions to make these problems visible.

We also discovered that catalogue filtering must be deterministic. An LLM should never recommend women’s clothing to a male customer simply because the description appears semantically relevant. Gender, size, category, tenant ownership and inventory constraints therefore run before probabilistic ranking.

Footwear introduced a separate routing challenge because it cannot be treated as a standard clothing try-on request. The system now selects the correct engine based on product category and verifies that the expected output was actually produced.

Maintaining a compelling live demo was challenging as well. Real API calls have variable latency, so we had to synchronize the narration with specific visual chapters while still allowing the live workflows enough time to finish.

Finally, multi-tenancy required careful attention. A useful recommendation is not enough if one retailer can accidentally access another retailer’s catalogue, customer memory or usage data.

Accomplishments that we're proud of

We are proud that TotalLook is more than an isolated virtual try-on screen. It demonstrates an end-to-end commercial architecture that retailers could integrate into their own customer experiences.

The platform combines skin-aware colour analysis, policy-based catalogue filtering, clothes try-on, footwear try-on, agent orchestration and human judgement in one traceable workflow.

We are particularly proud of:

• Keeping the retailer—not TotalLook—at the centre of the customer experience.
• Supporting multiple tenant catalogues through the same A2A gateway.
• Applying deterministic eligibility rules before AI ranking.
• Routing clothing and footwear to their correct YouCam engines.
• Providing traceability for every recommendation and try-on.
• Preserving service continuity through deterministic fallbacks.
• Escalating low-confidence decisions instead of hiding uncertainty.
• Demonstrating the platform through storefront, web-client and terminal-client perspectives.
• Supporting pseudonymous interactions that reduce unnecessary exposure of customer identity.

What we learned

We learned that personalization is most valuable when it is delivered as infrastructure, not as another destination the customer must discover and learn.

We also learned that generative AI should not control every decision. Language models are useful for interpretation, ranking and explanation, but catalogue eligibility, tenant isolation, sizing rules and policy enforcement need deterministic safeguards.

Virtual try-on quality depends on the entire input pipeline. Image accessibility, framing, format, product photography and category metadata are as important as the generation model itself.

Another important lesson was that confidence should affect workflow behaviour. A system becomes more trustworthy when it can distinguish between a recommendation that is safe to deliver automatically and one that deserves human review.

Most importantly, an impressive visual result is only one part of a production-ready fashion platform. Retailers also need privacy, observability, cost controls, explainability and reliable integration contracts.

What's next for TotalLook

The next step is to evolve TotalLook from a complete prototype into a retailer-ready platform.

We plan to add:

• More robust body-shape, fit and sizing analysis.
• Multi-item outfit composition and coordinated full-look generation.
• Additional garment and accessory categories.
• Catalogue ingestion from existing commerce platforms.
• Retailer analytics connecting recommendations to conversion and returns.
• Customer-controlled preference and wardrobe memory.
• Background processing and webhooks for long-running try-on jobs.
• Expanded consent, retention and data-deletion controls.
• More advanced stylist collaboration and approval workflows.
• Internationalized colour, sizing and accessibility support.
• Optional image-to-video generation for premium campaign and social-commerce experiences.

Our long-term vision is for any retailer to connect its storefront to TotalLook and immediately offer an intelligent, inclusive and auditable styling experience—while keeping its brand, catalogue and customer relationship entirely its own.

Built With

• postgresql
• qwen
• ruby
• ruby-on-rails