Client projectLiveArchitecture and design · AI Agents
OBJX Intelligence Multi-Agent Platform
Built by Van Data Team for OBJX Intelligence
Agent-first multi-tenant platform built around persistent memory, covering project management, financial operations, proposal generation, and a real-time dashboard.
- platform health score in production
- 94.7%
Measured outcomes
Outcome 01
94.7% platform health score in production
Outcome 02
Five specialised agents with MEM0 long-term memory
Outcome 03
Five-tier permission model with multi-tenant isolation
Challenge pressure
OBJX Design's AI platform needed agents that remembered prior work across sessions while staying isolated per tenant.
The situation
OBJX Design is an architecture and design practice in San Diego, and OBJX Intelligence is its AI platform, which also powers the practice's feasibility analysis, zoning review and development consulting. Agents that forget everything between sessions cannot run a programme of work. The suite spans programmes, projects, and tasks alongside financial operations, so the same agent has to pick up context it established days earlier, without ever seeing another tenant.
What we built
A five-agent architecture on LangGraph and LangChain over Claude Sonnet and GPT-4, with MEM0 holding long-term memory. The backend is FastAPI 0.104 on PostgreSQL 15, Redis, async SQLAlchemy 2.0, Alembic, and Celery; the frontend is React 18.3 and TypeScript 5.5 on Vite with shadcn/ui and TanStack Query v5. A five-tier permission system enforces multi-tenant isolation, and six enterprise integrations spanning Google Workspace, Slack, Teams, Salesforce, Monday.com, and Asana connect over OAuth2 and webhooks.
What changed
The platform reports a 94.7% health score in production, and OBJX now offers it to other professionals at meetobjx.com.
Constraints and decisions
Persistent memory and multi-tenancy pull against each other: memory is only useful if it survives sessions, and only safe if isolation holds at every tier of the permission model. Isolation is enforced in the data layer rather than in agent instructions, because an instruction is not a boundary.
Technology stack
Need a similar system?
If this case study maps to your current workflow, the next useful step is a scoped project conversation with constraints, timeline, and integration needs.
Related case studies
More work like this
Client projectLiveSmall business operations
A supervisor agent that runs 70+ business tools for small-business owners
70+
MCP tool integrations under one supervisor
Cybersecurity and compliance
A fine-tuned Mistral 7B behind an EU-sovereign compliance platform
424k
fine-tuning samples, plus 106k for validation



