I direct AI systemsthat buildproduction software.

Our team ran on Jira, Confluence, and five separate Apps Script and Google Sheets tools: project tracking, docs, internal and external time tracking, capacity and time off, an invoice portal, sprint summaries. Each one worked on its own. Together they fragmented every week, and clients had no window into any of it.

One platform for both audiences. The team runs delivery in it - sprint health, capacity, KPIs - and clients sign into the same product and see their own invoices, timesheets, and delivery status, scoped to their tenant and nothing else.

I defined the product, the tenancy and role model, and the portal boundary, then built it with Claude Code. What a client may see is a design decision enforced with row-level security, not a styling choice. Integrated with MAGNUS for the team-facing side.

In production for the OmniStreak team and its clients, with per-tenant scoping and audit logs. Flip the toggle above to move between the internal view and the client portal.

If your operations live across scattered spreadsheets and disconnected tools, this is what consolidation looks like when AI does the building and one person owns the outcome. The same approach applies to any internal tooling debt.

I watched the same QA breakdowns happen across dozens of client products, with teams stitching together five separate tools to cover one release cycle.

Eventually I had to build the tool that should have existed already. One hub, one login, and a family of focused apps under it: visual comparison against design, screen capture, speed audits, tracking checks, test case creation, team availability.

Built with Claude Code as the execution engine, with myself as the architect. A central API owns organizations, subscriptions, and entitlements; SSO carries one login across every app; each app checks access against the same source of truth.

Pre-launch. This simulation shows the hub and Visual Check, the app that compares a Figma design against the live page.

If you are weighing whether AI-assisted delivery can carry a full product from zero - architecture, multi-tenancy, billing, eleven applications - this platform is the answer I can show rather than argue.

A large production codebase, people who wanted AI assistance on it, and a hard constraint: AI could not be allowed anywhere near production, and client code could not be exposed to risk.

Instead of giving AI access, I built it a room to work in. A private mirror of the codebase - snapshots pulled in manually, scrubbed of secrets - where Claude does review, documentation, and ticket work without ever reaching a production repo.

I designed the roles and the refusal rules, and wrote them into the workspace. An admin gets the full surface. A stakeholder gets an assistant that answers scoped questions about the code and the tickets, and declines everything else - politely once, then silently. Output rules are enforced with hooks, and every session is logged.

Running inside my longest client engagement - 10 years and counting. Everything you type in this simulation runs against invented example content, not real code.

If your engineers want AI on the codebase and your risk answer so far has been no, this is the middle path: the team gets AI assistance, production and IP stay sealed, and every session is accountable. I have run it inside a real client relationship, not a lab.

Every operational question in a company lands on a person: HR policy, an invoice detail, an engineering convention, a legal basic. In a 20+ person remote team, that is a constant tax on exactly the people who have the least time to pay it.

When I see a process that could be automated or dramatically improved, I start planning a solution almost involuntarily. I built MAGNUS for my own team before most companies were thinking about internal AI infrastructure.

I designed the architecture and every agent's mandate, and built it with Claude Code. One Slack identity, 17 agents behind it; a lightweight router classifies each message and dispatches it to the right agent. Every agent runs on a system prompt I wrote and approved, with the same five guardrails ingrained in all of them.

In daily use inside OmniStreak across HR, ops, engineering, finance, legal, commerce, and delivery. The exchanges here are scripted replays of the kind of work it does; the real MAGNUS talks to my team, not to visitors.

If “we should have an internal AI assistant” keeps surfacing in your leadership meetings, this is what exists after that meeting when someone actually builds it: agents with mandates, guardrails, and an owner. I have done this once for my own company; doing it a second time for yours is a known path.