I build AI systems that
drive business outcomes.

Problem

A US-based AI sales company had prototypes for different sales automation workflows but no scalable architecture. The founder's vision was a templatized system where client configurations, prompts, and values could be organized centrally:enabling the agency to onboard new clients without rebuilding from scratch each time. When I joined, there was no production pipeline, no client management platform, and no way to scale beyond manual setup.

Decisions

The core architectural decision was separating strategic reasoning from natural language generation. AI models tend to conflate what to say with how to say it:producing messages that are either strategically sound but robotic, or natural-sounding but tactically weak. I split these into distinct agents:

I chose Supabase for the client architecture layer: storing all client configurations, prompt templates, and conversation state in a structured database with a web-based admin panel.

Outcome

The pipeline I designed processes 20,000+ conversations per month across active client accounts. The product it powers commands an $8-10K setup fee and $2K monthly retainer:positioning it as a premium offer in the Instagram coaching market rather than a commodity tool.

Building at this scale surfaced constraints that don't appear in smaller deployments: token cost management across thousands of daily calls, latency budgets that force trade-offs between model quality and response time, failure handling for multi-step agent chains where any single node can degrade, and conversation state management across sessions spanning days or weeks. The templatized architecture reduced client onboarding time by 40%, and a self-service admin panel enabled non-technical team members to configure new clients independently.

What I Learned

Separating concerns in AI systems works the same way it does in traditional software, but the payoff is even larger. When one agent handles reasoning and another handles language, you can debug each independently. If I were starting over today, I would build this in TypeScript rather than n8n. The visual workflow tool was right for the first iteration, but as the system grew, the ceiling became visible.

Problem

I wanted to grow on LinkedIn and spent most of 2024 trying every tactic available:courses, ebooks, templates. Nothing produced consistent results. I noticed a pattern from Instagram automation: mass engagement combined with selective exposure could accelerate follower growth dramatically. No tool existed that applied this approach to LinkedIn.

Decisions

I built the first version as a manual service, validating demand before investing in product development. The system used Selenium with Python for browser automation. Once followers started asking me to replicate my results, I invested in building a self-service SaaS portal. Growth came through organic product-led acquisition: I used my own growing LinkedIn presence (doubled from 2,000 to 4,000 followers in the first month) as proof-of-concept marketing.

Outcome

The product reached 70+ paying clients with minimal churn. I chose to sunset the product in April 2026. A deliberate product judgment call. The system depended on LinkedIn's tolerance of automation, and the risk of platform policy changes wiping out the entire business was high. Rather than wait, I redirected focus toward enterprise AI systems where the value creation is more durable.

What I Learned

The most important product skill I developed was knowing when to exit. LeadsAutopilot was profitable and growing, but the platform dependency risk meant the business had a ceiling that no amount of engineering could raise. It also validated the service-first, product-second approach: manual delivery to validate demand, then automation to scale.