AI Jade

Overview

AI Jade is an AI assistant layer that sits across a multi-plugin business platform. It helps users find features, analyse content and run multi-step workflows on their behalf. This engagement hardened and extended it. That meant faster responses and safer error handling. It also meant strict tenant and permission enforcement, cross-plugin workflow execution, AI content generation from images and meeting transcripts and real-time visual verification. It serves businesses whose teams work across several connected modules and want one assistant that respects who is allowed to do what.

Challenges

The assistant was slow. There were no baselines or monitoring to show where the time was going or to catch regressions. Several plugin endpoints were unreachable because their URL configuration had never been wired into the main application routing. Asynchronous views called synchronous methods without proper handling and behaviour was unstable as a result. When something failed users saw raw exceptions and tracebacks while the actual details went unlogged. That's the worst of both worlds. Public endpoints accepted unvalidated payloads. Project queries and object lookups were unfiltered. One customer's data could genuinely appear in another's session. Endpoints checked only that a user was logged in rather than whether they had permission for the specific feature they were invoking. Several headline capabilities weren't connected to real services at all. Image analysis, draft generation and visual verification were placeholders. Expensive AI and image operations had no rate limits so the platform was exposed to abuse and runaway costs. List endpoints returned hardcoded limits with no pagination.

Solution highlights

We started with measurement rather than guesswork. Profiling, baseline measurement and ongoing monitoring of key functions found the slow paths and now catch regressions. We fixed async and sync consistency across views and services. Database indexes were added across the query patterns that were actually hot. Error handling was centralised. Users see safe messages while full tracebacks, request IDs and retry context are logged internally. Circuit-breaker behaviour protects calls to external services. Input-validation serialisers were added to public endpoints. The plugin API URLs were registered so those endpoints became reachable at all. On security we built a permission service enforcing both plugin-level and feature-level access using the existing membership model. It applies across workflow, live analysis, image analysis and transcript services with explicit tenant filtering on project and object lookups. Then we built the capabilities that had only been stubs. Real AI image analysis extracts colour palettes, style and design elements. A transcript-processing workflow turns meeting recordings and inspiration images into structured draft moodboards, recipes and proposals through background processing. Live visual verification sessions compare real-world visuals against stored designs with real-time scores. We added cross-plugin workflow execution so a multi-step action runs as one orchestrated flow with permissions validated upfront and step-by-step progress tracked.

Key Features

• Performance profiling, baselines and ongoing monitoring with regression tracking
• Database indexes tuned to real query patterns across tenants, projects and sessions
• Centralised error handling with safe user messages, request IDs, retries and circuit breakers
• Input-validation serialisers on public endpoints
• Permission-aware plugin discovery and routing across all registered features
• Plugin-level and feature-level permission enforcement using the existing membership model
• Explicit multi-tenant isolation on project and object lookups
• Cross-plugin workflow execution creating projects, moodboards, recipes and proposals in one flow
• AI image analysis for colour palette, style and design element extraction
• Transcript processing that turns meetings and images into structured drafts
• Live visual verification comparing real visuals against stored moodboards and proposals
• Rate limiting, pagination with metadata and a test suite covering security and tenant isolation

Our Role & Approach

Our role was to take a promising but fragile assistant and make it safe enough to trust with real customer data and real actions. We sequenced the work by risk. Security and isolation came first because cross-tenant exposure cannot be allowed to reach production. Then reliability and error handling. Then performance. Feature completion came last. We put enforcement in shared services and decorators rather than in individual endpoints so permission checks can't be forgotten as the platform grows. Graceful degradation was a design principle. When an external vendor fails the assistant returns a service-unavailable state or a sensible fallback rather than a stack trace. A test suite covers tenant isolation, permissions, rate limiting, validation and error safety.

Technology Stack

Backend: Django REST Framework, django-tenants, Celery, serialisers, pagination, management commands. Frontend: React, React Router, WebSocket integration. AI and agents: Google Live API, Google Gemini Live API, OpenAI Vision API, Gemini Vision API, orchestrator agent, image generator, recipe generator, proposal agent. Integrations: moodboard, floral recipe builder, proposal/invoice, tenant CRM and live console plugins. Security: feature permission enforcement, tenant isolation, rate limiting, input validation, safe error handling.

Outcomes

The assistant responds faster and continuous monitoring catches regressions before users feel them. Customer data stays properly isolated. Tenant boundaries, plugin access and feature permissions are all enforced before any data is returned or any action is taken. Users are pointed only to the features they are entitled to use and the assistant can find and open those areas on their behalf. Approved multi-step actions run as a single orchestrated workflow rather than a series of disconnected manual operations. Meeting transcripts and inspiration images become structured draft moodboards, recipes and proposals automatically with a notification when they are ready. Design work can be checked in real time against stored references with comparison scores and feedback. Failures degrade gracefully through safe messages, retries and circuit-breaker fallbacks. Each operation is auditable through request IDs, workflow records and permission-denial logging. AI and image costs are controlled through rate limits. Pagination and indexed queries keep large datasets responsive.