Project Title:

AÉVRA — AI-Driven Horological Conservation Atelier & Dual Cockpit

The Hook (Opening line that grabs attention):

"Every mechanical watch has thousands of micro-components—and usually, a customer experience stuck in 2004. We rebuilt the horological conservation pipeline from scratch."

Overview & Story:

Most watch restoration platforms alienate customers with esoteric technical jargon and static forms. AÉVRA was created to preserve timepieces for what they mean, not just what they cost—combining editorial Swiss atelier aesthetics, real-time 3D motion, and AI-driven intake.

The Problem:

Customers with heirlooms or vintage pieces don't know reference numbers or movement calibres.

Atelier watchmakers spend hours manually triaging vague emails before diagnosing a single watch.

The Solution & Key Features:

Natural Language Horological Intake: Clients describe their watch issue in plain English (e.g., "losing time and scratched crystal"). The system automatically maps symptoms to mechanical calibres, delivering immediate clarity on services, timeline, and preliminary cost estimates.

Cinematic 3D Photographic Scroller: A high-fidelity scroll journey that physically disassembles and reassembles the movement's case, bezel, balance wheel, and gear train at 60fps.

The Conservator’s Owner Cockpit: A private, real-time management dashboard allowing atelier owners to view incoming requests, evaluate AI diagnostic breakdowns, and accept or decline repair commissions with instant, seamless switching between client and owner views.

Tech Stack & Craft:

Platform: Lovable

Frontend: React, TypeScript, Tailwind CSS

Motion & 3D: GSAP ScrollTrigger, Three.js / Canvas, Lenis Smooth Scroll

State Management: Zustand with local persistence

Live Project Link:

🔗 https://atelier-time-sync.lovable.app

Nice design

Complete stack RAG driven document intelligence assistant developed using FastAPI and React. Features multi-format ingestion support (PDF, DOCX, and scanned images through EasyOCR), with FAISS vector indexing, and 6 LLM driven modes operating on Groq’s Llama 3.3 70B model. Dockerized application via GitHub actions CI pipeline for automatic linting and build validation with sub-2 second response time in local tests. FAISS is chosen over Pinecone for quick in-memory computation without relying on cloud infrastructure, while Groq is selected over other options for quicker and cheaper inference than GPT-4 alternatives.

How do you persist the FAISS index across Docker restarts?