If you’ve been reading this newsletter for a while, you’ve watched me build something unusual.
I showed you the Personal OS.
I showed you the 33 automated tasks that run my consulting practice while I sleep.
I showed you what happens when the whole stack goes dark and you rebuild it in 72 hours.
I open-sourced a CRE acquisition framework - 54,000 lines of code, 31 agents, free on GitHub.
Every issue, I pulled back another layer. Architecture. Memory files. Agent orchestration. The self-improving system that gets better overnight.
But there’s something I haven’t told you yet.
While I was building all of that - and sharing it with you, issue by issue - someone else was building alongside me.
Same Obsidian + Claude Code stack. Same skill-based architecture. Same self-improving system.
He’s not a developer. He’s not a technologist. He’s not an engineer.
He’s a multifamily operator.
His name is Andrew Kuhn.
He’s the CEO of a Michigan-based multifamily company. President of the Apartment Association of Michigan. 20+ years in real estate.
His company fixes toilets and plows parking lots at 4 AM. His team sits with residents who are behind on rent and figures out a plan. Sunrise Communities provides housing for real families in real homes.
And in 61 days, he built an AI operating system that does this:
- 73 custom AI skills (if you’ve been reading along, you know what a skill is - a text file that gives AI a specific capability)
- 33 automated tasks running from 4:30 AM to 11 PM daily
- Nearly 50,000 files in a living knowledge vault
- 47,000+ node knowledge graph mapping relationships between owners, companies, brokers, properties, and deals
- 1,096 git commits - two-thirds generated by the system itself
- 5 AI team members with names, personalities, and defined roles
- Cost: ~$400/month in AI compute
Read those numbers again.
Then remember: this was built by someone who had never touched Obsidian, never written a PowerShell script, and never architected an AI system before January 16, 2026.
How It Started
Before the build, Andrew’s company ran the way every operator’s company runs.
Three email accounts checked manually. Deal tracking in spreadsheets that were always stale. Follow-ups falling through cracks because there was no system reminding anyone. Thirty minutes before every call trying to remember what was discussed last time. Underwriting deals over multiple days with manual Excel models.
The business ran. But it ran on Andrew’s memory, Andrew’s discipline, and Andrew’s calendar reminders.
He tried the off-the-shelf route first. Evaluated CRM platforms built for real estate. Explored AI tools and automation platforms. Even invested in companies building AI products for the industry.
Every single one hit the same wall.
He calls it the 60/40 Problem: commercial tools delivered about 60% of what he needed. The other 40% - the part that actually matters for how a specific company operates - required workarounds, compromises, or was missing entirely. No CRM understood the multifamily acquisitions workflow at the depth required.
No automation platform could connect relationship intelligence to database queries. And the companies building these tools were moving too slowly. By the time a feature shipped, the underlying AI had already passed it.
“I couldn’t find a tool that understood my business the way I understand my business. So I built one.”
The 61-Day Build
Here’s the compressed timeline. It reads like fiction. It isn’t.
Day 1 (January 16): First commit. Empty vault called “Andrew OS,” renamed to “Sunrise OS” that same afternoon.
Day 4: Imported tens of thousands of property files and owner profiles from commercial databases. The vault went from dozens of files to tens of thousands overnight.
Day 6: Evaluated a commercial CRM, studied its UI patterns, then made the call - “Obsidian + Claude Code is now the CRM.” No external CRM. Build it ourselves.
Days 9-11 (”The CRM Weekend”): Built 14 custom AI skills in 72 hours. Fixed 4,078 data quality issues. Merged 52 duplicate records. Ran overnight enrichment on 133 owner profiles using parallel AI agents.
Day 12: Completely redesigned the prospecting system. Threw away weeks of work because the new approach was better. Most people can’t do that. Andrew does it without flinching.
Day 20: The underwriting breakthrough. 8 complex deals analyzed in 20 minutes using parallel AI agents. Work that previously took days.
His reaction: “We just freed up three roles worth of manual work so the team could focus on residents and deals.”
Day 24: Hit the $200 monthly AI compute limit with 20 pending deals in the queue. Found workarounds. Separated automation compute from interactive sessions. Kept building.
Day 38 (”The Awakening”): The system started running autonomously. It began proposing improvements to its own instructions. Learning from its own operations. The first automated commit appeared at 8:00 PM. The system Andrew went to sleep with was better than the one he woke up to.
Day 61 (March 18): Current state. 73 skills. 33 automated tasks. Nearly 50,000 files. A knowledge graph with 47,000+ nodes and 55,000+ relationship edges. 55,000+ lines of automation code. The system improves itself every night.
What a Day Actually Looks Like
This is the part that matters. Not what the system is - what it does.
Before dawn (4:30 - 6:00 AM), the system runs a full sweep: vault integrity check across nearly 50,000 files. Broker listing sites scanned for new properties. Property management data synced - occupancy, delinquency, work orders. Cross-system signals analyzed.
By 6:30 AM, before Andrew touches his phone, a full briefing is waiting: emails triaged and prioritized, calendar prepped with context on every meeting, deal pipeline status, portfolio health snapshot. Plus a prioritized action list with full context on every deal and follow-up he owes.
Throughout the day, email gets triaged automatically. Meeting transcripts get pulled, analyzed, and intelligence gets routed to the right files. Research newsletters get converted into structured intelligence briefs.
By evening, the system scores Andrew’s day across 5 coaching frameworks, delivers targeted coaching questions, reads his responses the next morning, and adjusts its approach based on patterns.
Overnight, the system reviews the day’s work, proposes improvements to its own instructions, and auto-applies the safe ones.
“Today I drove four hours round trip to pick up my kids from my parents in Ohio. Before I left, Jarvis had prepped a full brief on every deal in my pipeline and every follow-up I owed. I made calls from the truck. I logged notes by voice. The system didn’t stop working because I wasn’t at my desk. That’s the point.”
The Numbers That Hit Different When You Have Context
If you’ve been following this newsletter, these will land harder than they would for a first-time reader. Because you’ve seen my versions of these same things.
You’ve seen my 33 scheduled tasks. Andrew built 33 of his own - different domain, same architecture, same self-healing infrastructure.
You’ve seen my skills. Andrew built 73
- 15 coaching personas plus 58 operational skills covering deal underwriting, prospecting, post-call coaching, market research, legal advisory, tax strategy, and content generation.
You’ve seen my automated commits. Andrew’s system generated two-thirds of its 1,096 commits. The system literally produces more work than the CEO does.
You’ve seen my vault. Andrew’s has nearly 50,000 files - every owner, every broker relationship, every deal, every call, every meeting. Logged, cross-linked, and queryable.
And the thing that stops people cold: he spent about $400 a month in AI compute.
“This isn’t a $200,000 hire. It’s not a $50,000 software platform. It’s $400 a month in compute. And it does work that wasn’t getting done at all before - not because we didn’t want to, but because there weren’t enough hours in the day.”
He Built a Team of AI Agents
The system didn’t stay personal. It expanded into the whole organization.
Andrew built 5 specialized AI agents for his team:
- Alice - Financial Analyst. Processes property-level P&L data, runs portfolio analytics, generates team dashboards.
- Lenny - General Counsel. Housing law, employment law, real estate operations legal questions across Michigan, Ohio, and Indiana. 30 reference topics.
- Tom - Tax Strategist. Entity structuring, depreciation analysis, 1031 exchanges, multi-state planning. Designs tax structures BEFORE deals close, not after.
- Sunny - General team assistant for day-to-day operational support.
- Oscar - Operations AI for daily scaling methodology.
The Walls
I’m not going to skip the hard parts. Because if I did, you wouldn’t believe the rest.
Authentication tokens expired overnight and silently killed 24 automated tasks. Andrew woke up to a dashboard that showed green when everything was broken.
Hit the monthly AI compute limit with 20 pending deals in the queue.
Discovered that cloud integration tools don’t work in automated processes. Built native API adapters from scratch in a single session.
An antivirus update crashed the entire system at 2 AM.
File sync conflicts corrupted data across three devices.
Each time, same response: fix it, learn from it, keep building.
“Nobody hands you this. There’s no ‘install AI’ button. We hit walls - authentication failures, architecture that didn’t scale, automations that broke at 3 AM. Every time, we fixed it and kept building. Not because we’re special - because we didn’t quit.”
The Method Works
Here’s why I’m telling you this story.
I’ve been the one on stage. For 10+ issues, I’ve been saying “look what I built.” And the skeptic in the back of the room has been thinking:
“Of course you built that. You’re the AI guy. What about the rest of us?”
Andrew answers that.
He ran both careers simultaneously for 14 years - medical device sales by day, building a real estate company on the side. That grind shaped the work ethic. But when it came to AI, he started from the same place everyone else does: knowing how to use ChatGPT and not much else.
The difference wasn’t technical ability. It was a decision.
“Avi didn’t build this for me. He taught me how to build it myself. That’s the difference between hiring a contractor and having a thought partner.”
We worked together for 40+ sessions between January and March. Most were 50-75 minute deep working sessions, not status calls. 2-3 per week during intensive build phases. I brought AI implementation expertise. Andrew brought 20 years of domain knowledge in multifamily operations. We were peer-to-peer from Day 1.
But the key word in that quote is “myself.” Andrew built this. I showed him the architecture, the patterns, the security considerations, the simplicity imperative.
He took all of it and built a system that’s different from mine in every way that matters - because it was built for his business, his workflows, his competitive advantages.
“We’re not a software company. We’re not a tech company. We provide housing for people. Real families, real homes. That’s exactly why this matters - if AI can change operations in a company that literally puts roofs over people’s heads, it can work anywhere.”
I built this for myself. I showed it to you issue by issue. I open-sourced the framework. And now, a multifamily apartment company CEO built his own version in 61 days for $400 a month in computing costs.
This isn’t about me anymore.
P.S. If this story made you rethink what’s possible for your business, forward it to someone who needs to see it. Forward it to the operator who’s been saying “AI isn’t ready for what we do.” It’s ready. And Andrew’s proof.
Andrew just started sharing his story publicly on LinkedIn - go find him and follow along. You’re going to want to see his perspective on all of this. He’s about to post his insights on his entire build, and you don’t want to miss it.