I made an AI "twin" of myself

I wanted to see: if I fed a system a structured framework about how I think, work, and make decisions — beyond the resume and titles — how would it respond?

That question became a build.

More than a resume

The first thing I figured out pretty quickly is that a resume is a terrible source of truth for who you actually are at work. It tells you what someone did. It doesn't tell you how they think, how they handle ambiguity, what they read, or how they explain themselves when no one is editing them.

So I started collecting a different kind of source material. Personality assessments. Books I've read. Hobbies. Project write-ups that captured reasoning, not just outcomes. FAQs. An AI guidance layer that shapes tone and framing. Markdown documents I can ingest, edit externally, and import back in.

Each source adds a different kind of signal. Structured data tells you what. Narrative content tells you how. Personal context tells you why. The blend is what makes the system feel like a person rather than a profile.

The voice interview

Here's the thing about written sources — they're all edited. Even when you're trying to be candid, you're still composing. Choosing words. Shaping a narrative.

So I tried something I didn't expect to work as well as it did. I had Claude interview me in voice mode. Unscripted. No chance to revise before answering.

The goal was to capture how I actually answer questions when I'm thinking out loud — the verbal reasoning, the qualifications, the way I circle back mid-sentence. That conversation became part of the knowledge base. There's a real gap between how you write yourself and how you actually explain yourself. This was an attempt to close it.

The build

For planning I used Perplexity. The actual product was built with Lovable and Supabase — lightweight enough to move fast, functional enough to be real.

The system has an admin console where I manage content across separate domains: Profile, Experience, Projects, AI Guidance, FAQs, Markdown import, and Chat Logs. That separation matters — it's not one giant prompt, it's a structured knowledge system where each section does a specific job.

On the features side: analytics to see what people actually ask, chat log tracking, a Calendly integration so someone can book time directly from a conversation, and embed support so ClaudIA can live anywhere via iframe.

There's also a human-in-the-loop layer. I can review real outputs, flag anything that's off, and promote corrections into the live system without touching the model itself. Lightweight governance — but it works.

Does it actually work?

I stress tested it with questions I hadn't planned for. Edge cases. Nuanced ones. The kind of thing where a static FAQ falls apart completely.

The replies were closer to my real answers than I expected. Not perfect — but close enough that the feedback loop felt worth running. When something was off, I adjusted. When something surprised me in a good way, I noted what made it work.

Oh — and the name. ClaudIA. The IA is intentional 😉 Inteligencia artificial in Spanish. A small bicultural wink from someone who's always moved between two languages, two markets, and two ways of seeing things.

She's live in her own sub-dommain ai.vclau.com if you want to meet her.