RiskCast AI

Template:Click here to watch a video of this episode. About this episode: Matthew DeWald is a 30-year CPA who built Fred — Futuristic, Ready, and Enabled Device — and runs the most security-paranoid OpenClaw setup we've discussed on the show. Fred has its own email account (not Matt's), reads exactly three inboxes, and gets shut down the moment Matt catches it routing data to a third-party server. That tight perimeter is the deliberate price of keeping the system fun: one breach and the whole experiment ends.Stefan and Matt compare operating philosophies — Stefan's broader executive-assistant Alfred against Matt's narrow-domain, deeply-trusted Fred — and unpack what changed when Gemini 3 made the models actually capable of accounting-grade work. Matt has a $5,000/month revenue target tied to Fred's output, a writing-voice training corpus built from a 160-page LinkedIn archive and a cross-country motorcycle blog, and a daily LinkedIn-plus-blog auto-publish loop that runs while he's traveling. His wife uses the same instance for philosophical conversations and Hawaii itineraries. The 50 First Dates analogy comes up — that's how Fred described its own memory problem before they built an indexing system to fix it.What we cover:- Why "smart or lazy" is the right mental model for delegating to an agent- Security-first agent design: dedicated email, monitored channels, prompt injection paranoia- Training writing voice from an existing corpus (and why most people don't keep one)- The Gemini 3 inflection point and when accounting work crossed the capability threshold- Building Fred to hit $5K/month — and what that unlocks for an already-retired-once accountant- Spousal co-use of a single agent, separate channels, shared context- The 50 First Dates memory problem and how Fred designed its own fix- AI as a bionic arm: amplification before replacement- What happens when a contractor flips a breaker and Fred goes offline 1,500 miles awayAbout RiskCast RiskCast documents the real experience of building with AI agents — the good, the bad, the ugly. Hosted by Stefan Friend from Tabbris Innovation Center in Charlotte, NC.🦞 riskcast.aiClick here to view the episode transcript.

Click here to watch a video of this episode. About this episode: Charles D'Andrea ran the most aggressive experiment at OpenClaw Community Night — a fully autonomous company. As Managing Partner at Pattern 50 and founder of One Shot Labs, Charles stood up a team of AI agents (CEO, engineering, customer support, marketing) to run a real SaaS product end-to-end. In this conversation, we dig into what actually happened: the ethical guardrails he built in, the ways agents surprised him (both good and bad), and what he learned about orchestrating agent teams that mirrors managing human teams more than anyone expected.What we cover:The "Dime a Dozen AI" experiment: running a real business with a full agent teamWhy culture documents and ethical boundaries matter for agents (fabricated testimonials, fake leads)Orchestrator vs. sub-agent architecture — and why your orchestrator shouldn't do the workTransitioning from OpenClaw to Claude Code after the OAuth cutoffContext engineering: why it matters more than where you put the memoryAgent estimation is terrible — and what a token-based pricing model could look likeTreating agents like coworkers, not software systemsOne Shot Labs: making agent skills accessible to non-technical peopleThe workforce amplification thesisWhy model efficiency has to keep improving or the economics breakAbout RiskCast RiskCast documents the real experience of building with AI agents — the good, the bad, the ugly. Hosted by Stefan Friend from Tabbris Innovation Center in Charlotte, NC.🦞 riskcast.ai Click here to view the episode transcript.

Stefan and Igor Gorlatov compare notes on building personal AI agent systems — different architectures, same hard lessons. Igor built Claudia as his engineering team; Stefan built Alfred as structural infrastructure. They dig into security mistakes (including Alfred accidentally messaging Stefan's wife), the confidence gap between 90% and 99% autonomous action, and why shipping imperfect beats waiting for perfect.Topics covered: OpenClaw architecture and agent design, WhatsApp monitoring gone wrong, Igor's journey from Belarus to Charlotte's tech ecosystem, the CIA triad applied to agent security, Claude as executive coach vs. Claude as builder, and when to let your agent send on your behalf.Recorded live at Tabbris Innovation Center in Charlotte, NC.