The problem
Property marketing often depends on a chain of specialists and disconnected tools. A realtor may need to clean up a room, correct light and weather, stage an empty interior, create a floor plan, and then hand the same media to an editor for video and social formats. Generative models can perform individual transformations, but a collection of model endpoints is not yet a product.
VisuGenie turns that fragmented process into a guided workspace. Users upload listing media once, choose a purpose-built task, monitor processing, and keep the finished image or video in the same project library. The platform owns the operational layer around AI: identity, team access, validation, model routing, credits, storage, status, and delivery.
My role
My work spans the full product boundary: React application flows, Django API design, asynchronous generation, provider integrations, billing and credit logic, production operations, and the public marketing experience. I translated fast-changing AI capabilities into stable, task-oriented interfaces that people can use without understanding prompts, model versions, or provider-specific parameters.
What I built
01. A focused product over a growing model catalog
The interface is organized around real jobs—declutter a room, add furniture, improve lighting, renovate an exterior, render a drawing, build a clip—not around provider model names. A task registry and schema-based contracts define required inputs, choices, prompt strategies, pricing, and the model that should execute each operation.
02. A two-repository production architecture
The end-user experience is a React 19 and TypeScript SPA using TanStack Query, Zustand, React Router, and Tailwind. A separate Django 5.2 REST backend owns users, teams, projects, media, task configuration, OpenAPI endpoints, and administration. This separation keeps the customer experience responsive while the server remains the authority for business rules and generation state.
03. Durable asynchronous AI workflows
Generation requests are validated and charged before Celery dispatches them to Replicate or FAL. PostgreSQL stores durable job and asset state, Redis backs the queue, provider webhooks reconcile completion, and generated files move into S3. Users can leave a page and return to the same project instead of depending on a browser request or temporary provider URL.
04. The commercial and operational platform around inference
Stripe subscriptions, team wallets, credit transactions, feature gates, refunds, and reconciliation make model usage accountable. Headless authentication supports email, Google OAuth, MFA, and WebAuthn. Wagtail content, localized interfaces, staff analytics, support tools, and configurable administration let a small team operate the product without shipping code for every content or model change.
From one upload to reusable listing media
- 01
Upload a property photo or select an asset already stored in a team project.
- 02
Choose a task; the UI collects only the inputs and options relevant to that result.
- 03
The API validates access, wallet balance, parameters, model availability, and task policy.
- 04
Celery submits the job, providers run the AI workload, and callbacks update durable state.
- 05
Post-processing applies the required crop, resize, watermark, or overlay before S3 delivery.
- 06
The result returns to the project library and can become the source for another image or video step.
Key engineering decisions
Task contracts as configuration
Shared task keys and input schemas keep frontend guidance, backend validation, prompts, pricing, and model routing aligned.
Provider independence at the product layer
Users select an outcome while the backend resolves the current model and provider, making model upgrades an operational change.
Generated media as first-class assets
Outputs live in projects with persistent lineage and storage, so results can be reviewed, reused, and chained into later steps.
Credits linked to actual work
Wallet transactions and asset charges preserve an auditable connection between a generation, its cost, and any correction or refund.
Result and reflection
VisuGenie is a live, multilingual product that unifies a broad real-estate media toolkit: photo correction, staging, rendering, renovation concepts, floor plans, short-form effects, and assembled video tours. The important result is not access to any single model; it is a product system in which those capabilities can change without breaking how customers organize and finish their work.
The project reinforced a practical lesson about production AI: the model call is often the smallest part of the system. Clear task design, durable state, billing correctness, media lineage, and operational controls are what turn a compelling demo into a service people can return to.