Relja B.

I build AI products that run in production: agents that take real actions with a person approving the risky ones, LLM features inside SaaS and iOS apps, and the full-stack backend behind them. Most AI demos break the first time they touch real data, real money or a real customer. I build the parts that make them safe to ship: tools with strict permissions, human approvals, audit logs, evals on real cases, cost limits and fallbacks between models. What I've built - Gatehouse (open source): an MCP server where agents read and propose freely, but a refund, a discount or an email to a customer only happens after a named person signs off on that exact action. 13 of 13 red-team attacks blocked, 72 tests. - DeepLead: a sales intelligence platform on the Claude Agent SDK with 18 agent modules, cost controls, CRM writeback, Stripe billing and 40 automated tests and evals. - LaunchPilot: an AI marketing platform in Next.js and Supabase with 86 API routes and 15 scheduled jobs, routing between Claude, OpenAI, Gemini and Grok, and publishing to six social networks through their official APIs. - SupportAgent: answers customer email for several products and hands refunds and edge cases to a human. - Vael: AI wardrobe and virtual try-on. One Supabase backend serves a SwiftUI app, a Next.js web app and a Chrome extension. - PayControl: a FastAPI fintech backend with 50 routes, Plaid bank sync that stays correct through retried and duplicate webhooks, and 36 tests. Full-stack and iOS I ship the whole product, not just the model call: Next.js and TypeScript, Node, Python/FastAPI, Supabase/Postgres and Firebase, Stripe, and native iOS in Swift and SwiftUI. 11 of my own apps are live on the App Store, and AirPosture has 2,000+ users and paying subscribers. On Upwork I'm Top Rated with 100% Job Success, and every client has left 5 stars. Clients here have hired me for AI chat and voice-first AI apps as well as native iOS work. How I work I build with Claude Code every day inside my own agent setup: agents work in parallel on separate copies of the code, and nothing counts as done until a second agent has checked it and I have reviewed the diff. You get a team's speed with one senior engineer accountable, working builds often, and the code and accounts in your name from day one. Based in Belgrade (CET), working with US and EU clients, full-time or part-time.