AI Ekip.

Overview

AI Ekip is a Canadian AI SaaS with four specialized agents: Resume Builder, Health Advisor, Crypto Analysis, and General Assistant. Subscriptions ran at $29 and $49 per month. I owned the co-pilot product end to end for 18 months as the sole engineer on that surface.

Problem

The product had to go from early concept to production AI co-pilot. The architecture required real-time streaming, multimodal input, encrypted communication, and a subscription billing layer. I was the only engineer responsible for the co-pilot, which meant ownership across the full stack.

Constraints

• Backend APIs were not publicly documented. Reverse engineering was required to build reliable integrations.
• The platform had to work as both a Chrome extension and an Electron desktop app from one shared codebase.
• Content analysis that took 4 hours of manual work needed to drop to under an hour without losing accuracy.
• Test coverage had to be high from the start, not retrofitted. Real paying users were on the line.

Decisions

• Build the co-pilot architecture before building features. SSE streaming, dual-backend encrypted communication, and the multimodal flow as foundational decisions before any agent UI.
• Reverse-engineer the undocumented APIs and build a stable abstraction layer on top, so upstream changes do not break the rest of the codebase.
• Share one codebase across Chrome extension and Electron desktop. Two clients, one product.
• Treat 95% test coverage as a working tool, not a vanity metric. The only way to refactor without breaking the agents users depend on.

What I built

• Chrome extension and Electron desktop app from scratch, single shared codebase
• Real-time SSE streaming with encrypted dual-backend architecture
• Multimodal user flow: image capture, chat interface, streaming AI responses
• 4 specialized AI agents with distinct personas, tools, and response patterns
• AI-powered content analysis: manual review reduced from 4 hours to 45 minutes (75% reduction)
• Microservices architecture handling 18,000+ daily requests, scaled to 100,000+
• Subscription billing at $29 and $49 per month tiers
• 95% test coverage across the engagement

Result

• 3,000+ users on the platform
• 18K daily requests at launch, scaled to 100K+ as the platform grew
• 75% reduction in manual content review time (4 hours to 45 minutes)
• 95% test coverage
• 331 commits and 586 passing tests on the core co-pilot in under 4 months

Reflection

Building a production AI system solo teaches you that most of the work is not in the AI calls. It is in the reliability layer around them. Handling failures gracefully, making streaming feel smooth, and keeping the test suite honest so a refactor does not silently break an agent. The 95% coverage was not a vanity metric. It was the only way I could move fast without breaking things users depended on.

What my co-builder said