Craig Campbell

When I joined EV Life, the codebase had no clear owner, no documentation standards, and no defined architecture. Engineers were writing code that worked, but nobody could explain why decisions had been made or where the service boundaries were.

For a startup trying to move fast, this wasn't just technical debt — it was a compounding drag on every sprint. New engineers took weeks to get productive. Debugging meant archaeology. Every deploy carried risk that nobody could quantify.

I started by listening: understanding what the team knew that wasn't written down, where the real fragility was, and which parts of the system were genuinely at risk versus just undocumented. From there I introduced clear architecture docs, defined ownership boundaries for each service, and established standards for how new code should be structured and reviewed.

Critically, I did this incrementally — adding structure where it reduced pain, rather than a big-bang rewrite that would have stalled product work. In parallel, I architected AWS infrastructure using CDK with ECS/Fargate, designed and built a RAG-based AI system to automate EV incentive monitoring, and built a loan origination system that reduced per-application processing time from 6 hours to 30 minutes (~80% reduction) for a 5-person loan ops team processing 50+ applications/week — integrating Plaid, Kelly Blue Book, and multiple financial APIs.

The RAG system integrated live web scraping via Firecrawl, internal PostgreSQL context, and Anthropic LLM analysis — improving incentive-change detection accuracy from 10% to 90% while reducing co-founder manual research time from 8 hours/week to 15 minutes. It now monitors 443 incentive programs, 789 data sources, and 46 states, detecting ~30 changes per week.

Alongside building AI systems, I led the team's adoption of AI-assisted development practices — establishing guidelines that balanced delivery velocity with code quality, security, and data privacy. This included team guidelines for agentic workflows, MCP server integrations (Context7, Sentry), and prompt quality standards that improved LLM output accuracy and reduced development rework.

These guardrails weren't restrictions — they were what made the team confident enough to move fast. We also built SOC 2 Type 2 readiness in parallel, developing 25+ security policies and maintaining 200+ controls covering security, compliance, data privacy, and AI tooling risk — bringing the platform to full audit readiness within 2 months.