I built a personal AI assistant. Here's what a year of daily use taught me.

Late 2024, I got tired of re-explaining myself to AI assistants.

Every morning I'd open a new Claude session, paste in my calendar, remind it of my current projects, and hope it remembered the context I'd built up yesterday. It never did. Each session started from zero.

So I built something different.

What I actually built

I call it Cass — the version I run for myself. It's powered by Claude under the hood, accessible via SMS, and it maintains memory about my life and work across every session.

The key difference from ChatGPT, Claude.ai, or any other assistant product: no app, persistent context, and it reads my actual calendar.

Every morning at 7:30 AM, I get a text. Here's roughly what it looks like:

Good morning. You have a 10 AM call with the AGJ team, a 2 PM with a vendor, and the Frogger deploy needs to happen before end of day — you flagged it yesterday. Amanda's birthday is Thursday. Weather: high of 84, 20% chance of afternoon storms. Heads up: the $400 Hover renewal hits your card today.

No app to open. No prompt to write. It just shows up in iMessage like a message from a person who knows my life.

What makes it different

Access. Every AI assistant assumes you're at a keyboard ready to context-switch. I'm usually in the middle of something. SMS means I can ask "what's on my calendar Thursday?" while I'm in the car, or dictate a quick task while I'm making coffee. The access pattern changes how much you actually use it.

Memory. The assistant knows my company, my team, my current projects, and my preferences — because I've told it all of that, and it remembered. The first session is as informed as the hundredth. I'm not re-explaining the difference between my AGJ work and my personal projects every morning.

Integration. It reads my Google Calendar. Not "you can share a screenshot of your calendar with it" — it actually calls the API and pulls my schedule. When I ask what I have Thursday, it knows.

No hallucination guard rails. Most AI products are tuned to hedge constantly. Mine isn't. It's direct, it tells me when it doesn't know something, and it doesn't pad its answers with disclaimers. I briefed it on how I like to communicate, and it adapts.

What doesn't work (yet)

I want to be honest here, because the pitch for AI assistants is usually oversold.

It doesn't replace judgment. It can draft an email, but I'm still the one deciding whether to send it. It can surface a calendar conflict, but the decision about what to reschedule is mine. It's a very good first pass on almost everything, not a decision-maker.

Email read/draft is useful, not magic. I can ask it to summarize recent emails from a specific person or draft a reply in a particular tone. But I have to paste the email into the conversation — it doesn't proactively scan my inbox. That integration exists and I can build it, but proactive email monitoring has failure modes I'm not comfortable with for a first release.

Overnight follow-up is unreliable. If something important happened at 11 PM, I'll know at 7:30 AM. That's usually fine. If you have an overnight business that needs immediate escalation, the current system isn't wired for that (though it can be).

Context drift. Over a few months, the memory gets slightly stale. I do a quarterly "memory refresh" — go through what the assistant knows about my current priorities, update what's changed. It takes about 20 minutes and I do it while drinking coffee. Not a burden, but it's not zero maintenance.

Why I'm sharing it now

A few things clicked in the last month:

1. I realized I've been using this daily for over a year without getting tired of it. Most productivity tools I build don't survive contact with my actual life this long.

1. Two people in my network separately asked me "what's that thing you use that sends you a morning briefing?" after I mentioned it offhand.

1. I've gotten the setup down to about 3-4 hours per client. It's repeatable.

So I'm taking three pilot clients this summer.

What you get: a dedicated AI named Evangeline (clients don't interact with "Cass" — they get their own instance, briefed on their context, with their own Twilio number). SMS-accessible, remembers your world, reads your Google or Outlook calendar, sends a morning briefing at whatever time you want.

Tiers: $500/mo for the basic SMS + memory setup (Scout), $1,200/mo with calendar and email integration (Director), $2,000/mo with CRM integration and proactive alerts (Principal).

First 30 days are free. In exchange for 20 minutes of honest feedback after the trial.

If this sounds like something that would actually fit your workflow — not because AI is exciting, but because the access pattern of SMS-first with persistent memory is different from every other tool you've tried — reach out.

I have three slots. I'll be picky about the fit.

The actual technology, briefly

It's not magic. It's a Twilio phone number routed to AWS Lambda running Claude (Haiku for quick tasks, Sonnet for research and drafting). Memory lives in DynamoDB. Calendar access is Google Calendar API or Microsoft Graph with delegated OAuth. Morning briefings are EventBridge cron jobs that fire at the client's local time.

Per-client infrastructure cost: $15-30/month in AWS + Claude API costs. The rest is Chester's time (setup: 3-4 hours; steady state: <2 hours/month).

The interesting part isn't the infrastructure — it's the prompt engineering and memory management. What should the assistant know by default? How do you handle topics the client doesn't want to discuss? How do you write a morning briefing that's genuinely useful rather than a laundry list? I've iterated on those questions for a year on myself.

That's the moat. Anyone can spin up a Twilio webhook. Not everyone has 12 months of morning briefings to learn from.

If you're curious and want to see what the intake process looks like: pre-briefing form.

If you know someone who would be a better fit than you: forward this to them.

Three slots. Summer 2026.