Try-on Assistant
AI-powered backend system enabling virtual try-on (VTO) filters with conversational product recommendations.
Client:
Lashify, USA
Project Overview
MaquillAR, led by founder and creative director Sofi Chernyak, sought to expand their existing virtual try-on experience by adding intelligence and interactivity. The goal was to move beyond static filters and integrate a backend AI system that supports both real-time VTO and a conversational assistant capable of recommending products. Vadim Hursevich spearheaded the technical development, focusing on backend architecture, AI integration, and scalable deployment.
Challenge
The client’s existing system was built primarily on the Snapchat SDK, with no backend execution capability. This created several challenges:
Lack of Backend Logic: The VTO system could not process real-time product data or communicate dynamically with users.
Product Database Integration: There was no structured way to connect product data to the recommendation engine.
Complex AI Guardrails: The conversational assistant required strict guardrails to avoid unsafe or irrelevant responses (e.g., discussions about plastic surgery).
Performance & Latency: Ensuring fast responses (500ms–2.5s for text; 800ms–3s for voice) while maintaining accuracy and reliability posed a significant technical hurdle.
Ownership & Licensing: Negotiations around code reuse and intellectual property had to be carefully balanced between developer and client interests.
Tech Stack
Cloud: Google Cloud Platform (E2 instance, T4 GPU), with future migration potential to Azure
Databases: Vector DBs (Qdrant, Milvus) for embeddings and product search
Backend: Python, Node.js, REST APIs for SDK communication
AI / ML: Embedding models for product search and recommendations, GPT-based conversational assistant with prompt engineering and hallucination reduction
SDK Integration: Snapchat SDK for virtual try-on filters
Solution
The system was designed in phases to ensure iterative delivery and flexibility:
Classification of SDK Filters
Integration with the Snapchat SDK to categorize and optimize VTO filters, ensuring compatibility and extensibility.Conversational AI Assistant
A backend-driven chatbot providing tailored product recommendations while respecting safety guardrails and maintaining conversational relevance.Live Mode Integration
Backend APIs were built to support real-time interactions, bridging the gap between the client’s website and SDK-based experiences.Embedding Search System
Development of a vector-based search engine that uses product embeddings for personalized recommendations. This enables users to discover products dynamically within the VTO experience.
The outcome is a scalable AI-driven backend system that transforms MaquillAR’s virtual try-on into an intelligent, interactive, and product-aware solution, ready for continuous growth and brand expansion.
