George B.

I design and run AI agent systems in production. I co-founded Mojar AI, where I own the platform architecture and the agent runtime: tool approval rules, step and cost budgets, researcher and writer sub-agents, resumable runs that survive a dropped connection, and model routing across Anthropic, OpenAI and Google with per-turn cost accounting. The stack is Next.js, NestJS, Supabase Postgres with row-level security, BullMQ and the Vercel AI SDK, with a Model Context Protocol layer for knowledge-base tools and a CRM connector where external writes are denied by default. Recent client work: for a CRM company I delivered an agent harness architecture document and a working proof of concept in their TypeScript, NestJS, Supabase and BullMQ stack in one day, covering the execution loop, a Postgres-backed state machine, rep approval gates and the audit trail. The client wrote that it "looks like more than a proof of concept". What I am good at Agent harness architecture: tool definitions, execution loop, state, context management for long and multi-session tasks, human-in-the-loop checkpoints, failure modes. Shipped, not demoed. Supabase and Postgres security: row-level security as the authorisation model, grants and function hardening, missing indexes, and a CI check that blocks any migration that weakens access rules. At Mojar AI I led a full security audit of the product and remediated every finding as a separate reviewable change. Multi-tenant SaaS on Next.js and NestJS: as Chief Innovation Officer at Positive Action (2021-2026) I designed its platform: a federated GraphQL gateway over NestJS services for organisations, roles, seat licensing and certificates, a self-hosted Authentik single sign-on with OAuth2/OIDC for every application, and over 100 production releases of the core service. It ran on Kubernetes with GitOps. Payments and commerce: I built a multi-tenant ordering and payments platform alone on Next.js, Supabase and Stripe Connect, with per-organisation connected accounts, idempotent webhooks and full order lifecycles. How I work I read your brief and answer the question you actually asked. You get a deliverable with a date, not a list of hours: an architecture decision document, a prioritised findings report, a runnable proof of concept, or a reviewable set of pull requests. I write plain English, record decisions with their reasoning so the knowledge stays with your company, and ship in small steps you can review. Twenty years in software, the last ten in TypeScript and Node.js. Based in Romania; my working day overlaps US East until early afternoon and all of Europe. Good fits: agent harness design or review, taking an AI prototype to production, Supabase security audits, architecture review of a Next.js or NestJS SaaS, fractional architecture work for a small team. Not a fit: Python-only ML roles, mobile apps, Rails or ColdFusion codebases, jobs under $60/h. See Mojar AI running at mojar.ai.