About AI Frankly

AI, frankly told. Stress-testing the tools, the plumbing, and the messy reality of the AI stack.

Who writes this?

I'm Frank Cabrera. By day, I work in enterprise tech. For the last 20 years, my job has been figuring out why the expensive tools companies buy rarely work as advertised. I've seen enough hype cycles to know the difference between a glossy demo and something that survives IT, legal, and procurement at the same time.

AI Frankly is where I take that skepticism and apply it to AI, tech, gaming, and whatever I decide to break next.

I don't answer to a director. I don't have story quotas, tight deadlines, or a content void to fill. My only driver is my own curiosity about how things work behind the scenes.

Think of this as 'Behind the Music' for the AI industry. While everyone else is watching the glossy music video (the hype), I'm showing you the studio sessions, the plumbing, the orchestration, and the messy reality of the production.

What I'm Building

Project FrankOS is a publicly documented local AI executive assistant I'm building in the open. Every agent becomes a Substack post. The code is free. Current soldiers: The Fetcher. Next up: Calendar, Coder, Orchestrator. Follow the build on Substack.

Find Me

What Didn't Work (And Why I'm Telling You)

Ollama + LLaVA for document analysis

Spent a weekend trying to get LLaVA running locally for invoice parsing. It could describe a photo of a cat but couldn't reliably read a two-column PDF. The model hallucinated line items that didn't exist. I went back to Claude with a clean room setup. Sometimes local isn't the answer.

The "AI email triage" workflow

Built an agent chain that was supposed to read my inbox, classify urgency, and draft replies. It worked perfectly in testing. In production, it flagged a routine vendor renewal as "critical legal action required" and drafted a response that nearly ended a partnership. Killed it the same day. Automation without guardrails is just faster mistakes.

Microsoft Copilot's "enterprise-ready" claim

Tested Copilot across the Office suite for three weeks. Word integration was decent. Excel was inconsistent. Teams summaries missed action items half the time. The gap between the keynote demo and what actually ships is wide enough to drive a truck through. Enterprise-ready is doing a lot of heavy lifting in that marketing copy.

Need a sanity check?

Twenty years of enterprise IT. If you're trying to figure out where AI fits in your stack and need an honest read, book an hour.

Get in touch