work / case
A company-wide agentic AI platform
Sole owner of the secure, highly available AI platform 60+ staff depend on — rebuilt from a single-user open-source agent into a multi-tenant system with isolated per-user workspaces.
Problem
Off-the-shelf agentic coding tools are built for one developer on one laptop. The organisation needed something the whole company could lean on: every staff member with their own AI agent for personal assistance, automation, and hands-on agentic development — without sharing a workspace, leaking each other’s code or files, or fighting over a single set of credentials. The most capable open-source harness available assumed a single trusted user with full host access. Dropping 60+ people into that model would have been a security and reliability disaster: no tenant isolation, no high availability, and one shared blast radius.
Approach
I forked the single-user harness and rebuilt it from the ground up as a secure, multi-tenant enterprise platform, running on a production Kubernetes cluster I also operate.
- Per-user isolation. Each user gets an isolated workspace pod with its own compute, storage, and identity — the core architectural change that turned a single-user tool into a multi-tenant system. RBAC and NetworkPolicy enforce boundaries so one tenant can never reach another’s workspace.
- Hardening. Closed cross-user data-exposure paths (diff/proxy leakage), added token redaction for scheduled jobs, stripped auth cookies at the application proxy, and validated user-supplied volume overrides to prevent escape.
- Stateful, durable backend. Session, lease, and pub/sub state on Redis Sentinel; persistent state and metrics on PostgreSQL/TimescaleDB; shared project storage migrated to a CephFS-backed volume with snapshot and disaster-recovery workflows.
- High availability. Designed the control plane and storage so the platform survives node loss, because the entire company now depends on it being up.
- Sole ownership. I am the only engineer on the company deployment — architecture, implementation (a large Python codebase), security, operations, and the weekly product demos to leadership that drive the roadmap.
Outcome
The platform is now critical internal infrastructure used by 60+ staff as their AI assistant and agentic development environment. Each person works in an isolated, individually-resourced workspace; tenants are cryptographically and network-separated; and the service runs highly available with snapshot/DR coverage on shared storage. What started as a single-user tool became a system the whole company runs on — owned, operated, and hardened end to end by one engineer.