WanderCode logoWanderCode

AI Jade — Hardened, Permission-Aware AI Assistant for Cross-Plugin Creative Workflows

We turned a slow, fragile AI assistant into a fast, secure, multi-tenant platform that navigates plugins, orchestrates cross-plugin workflows, and generates creative drafts from images and transcripts.

AI Jade — Hardened, Permission-Aware AI Assistant for Cross-Plugin Creative Workflows

Overview

We partnered with AI Jade to harden and accelerate their AI assistant, enforcing strict tenant and permission boundaries while adding plugin navigation, cross-plugin workflow orchestration, AI content generation from images and transcripts, and real-time visual verification. The result is a faster, safer, more capable assistant that users and developers can trust.

Challenges

  • AI responses were slow, with no baselines or monitoring to pinpoint where time was being lost or to catch regressions.
  • Plugin endpoints were unreachable because their URL configuration was never wired into the main application routing.
  • Async views were calling synchronous methods without proper wrapping, creating inconsistent and unstable behavior.
  • Raw exceptions and tracebacks were exposed to users instead of safe messages, while internal failures went unlogged.
  • Public endpoints accepted unvalidated request payloads, leaving the system open to malformed and unsafe input.
  • Project queries and object lookups were unfiltered, risking cross-tenant data exposure between customers.
  • Endpoints checked only for authentication, not feature-level permissions, so users could reach actions they should not access.
  • Image analysis, draft generation, and visual verification were placeholders that never connected to real services or plugin APIs.
  • Expensive AI and image operations had no rate limits, exposing the platform to abuse and runaway costs, and list endpoints returned hardcoded limits with no pagination.

Solution highlights

  • Built performance profiling, baseline measurement, and ongoing monitoring of key functions to surface slow paths and track regressions, and added async/sync consistency fixes across views and service calls.
  • Added database performance indexes across tenant, status, project, sender, session, interaction, and live-analysis query patterns to speed up frequently filtered and ordered fields.
  • Introduced centralized error handling with safe user messages, internal traceback logging, request IDs, retry logic, and circuit-breaker behavior for external services.
  • Added input-validation serializers for email analysis, response generation, and other public endpoints, and registered plugin API URLs so endpoints became reachable through the main application.
  • Built a permission-aware plugin discovery and navigation system, backed by a navigation map model and management command capturing plugins, features, routes, endpoints, permissions, and available actions.
  • Implemented a permission service and decorators enforcing plugin-level and feature-level access across workflow, live analysis, image analysis, and transcript analysis services using the existing membership model.
  • Delivered cross-plugin workflow execution that creates projects, moodboards, recipes, and proposals in a single orchestrated flow, with wrapper services, status APIs, upfront permission validation, and step progress tracking.
  • Added AI image analysis for color-palette extraction, style analysis, and design-element extraction, plus a transcript-processing workflow that turns meeting transcripts and inspiration images into structured moodboard, floral-recipe, and proposal drafts via Celery async processing.
  • Shipped Google Live visual verification sessions and a frontend widget that compares live visuals against stored moodboards, recipes, and proposals with real-time scores, alongside rate limiting, pagination with page metadata, and a test suite covering security, validation, and tenant isolation.

Outcomes

  • The assistant responds faster thanks to profiling, indexing, async/sync cleanup, and continuous monitoring that catches regressions before users feel them.
  • Customer data stays isolated, with tenant boundaries, plugin access, and feature permissions enforced before any data is returned or any action is executed.
  • Users are guided only to features they are permitted to use, and the assistant can discover and navigate accessible areas on their behalf.
  • Approved multi-step actions now run as a single orchestrated cross-plugin workflow instead of disconnected, manual plugin operations.
  • Meeting transcripts and inspiration images are converted automatically into structured moodboard, recipe, and proposal drafts, with notifications when drafts are ready.
  • Design work can be validated in real time by comparing live visuals against stored moodboard, recipe, and proposal data with comparison scores and feedback.
  • Failures degrade gracefully through safe error messages, retries, service-unavailable states, and circuit-breaker fallbacks for external services.
  • Operations are auditable end to end via request IDs, workflow records, permission-denial logging, and stored analysis records.
  • AI and image costs are controlled through rate limits and request limits, while pagination and indexed queries keep large-data handling responsive, and a test suite gives developers confidence in security and isolation.

Gallery

AI Jade — Hardened, Permission-Aware AI Assistant for Cross-Plugin Creative Workflows screenshot 1
AI Jade — Hardened, Permission-Aware AI Assistant for Cross-Plugin Creative Workflows screenshot 2
AI Jade — Hardened, Permission-Aware AI Assistant for Cross-Plugin Creative Workflows screenshot 3