AIMUG is the Austin LangChain AI Middleware Users Group — a 1,700-member community for AI builders in Austin, Texas, founded by Colin McNamara. Monthly mixers run at ACC RGC 3000 with lightning talks, live demos, and an after-party at The Tavern. This is my recap of the May 6, 2026 AIMUG Monthly Mixer & Showcase — my first one.
I went to my first AIMUG tonight (Meetup page). Walked in cold. Walked out vibrating.
Then I sat on my couch, opened my notes app, and stared at this:
Lanai. Austin Collin cases Soschaba - German dude - hand tracking no chat app Nate - was walking agent orchestrator / subagents - part of software vs tech stack specialist (db, ledger, api, ect) Jordan- shouted out during beginning of automated publishing software openclaw
That’s the first six lines of my notes from the meetup. That’s what I came home with. Half typos, half stream-of-consciousness, one of them is just the word “openclaw” floating in space like a haiku.
So I opened Claude Code, hit /pa, and pasted the whole thing in.
What follows is the decode.
”Collin cases”
Let me translate live.
“Collin” — that’s Colin McNamara. I spelled his name wrong. He’s also the founder of AIMUG. The 1,700-member meetup I just walked into for the first time. The host. I had no idea.
“cases” — that’s Acquit.ai, the litigation-intelligence platform he’s building, which models opposing parties as cognitive agents over a corpus of court cases. 91% accuracy across 201 tracked predictions, by his published numbers. I caught a fragment of him explaining it and my thumbs typed “cases” before my brain finished the sentence.
The man’s resume reads like a trick: architected the world’s largest CDN at Apple iCloud in six months, $113M project. Senior Principal at Oracle running 4 global services across 64 data center networks. Built a $180M/yr line of business at NexusIS that got acquired by Dimension Data. USMC veteran. KeyBanc Mosaic Advisory.
Then he told me about his safety thing. (My notes do not say “safety thing.” My notes say nothing about it. I just remembered after the fact.) He’s extending FMEA — the framework you use in nuclear, aerospace, and food safety — into a risk-priority model for AI decision systems, where safety lives inside the objective function instead of as a downstream checklist. The argument is laid out on his site: Safety Should Be a Constraint, Not a Checklist. The implementation lives inside Acquit.ai.
I sat down on my couch, sent him a LinkedIn connection request without a note, and started reading.
Highest-leverage connection of the night by a country mile, and “Collin cases” is what my phone gave me to remember it.
”Jordan- shouted out during beginning of automated publishing software”
Translation: someone said “Jordan Hill” during the night’s intro and my notes app got into bed. Beginning of automated publishing software is, on inspection, what I typed when Jordan — founder of Organized AI, runs Clawstin (the Austin OpenClaw meetup) — was being introduced.
His talk: “Store · Work · Export — Building Agent Runtimes on Cloudflare.”
I have these direct quotes in my notes from his talk:
Build guides wikis architecture - FOR MY OWN PROJECTS! but be careful… you could obliterate your context if it’s too big. His example: ORGANIZED AI — Index of Everything Deployed
That’s pattern #1. Every system Jordan ships gets one self-contained HTML page documenting its architecture, components, and links. You drop the URL in front of any new collaborator, any new AI agent, any new investor — and they have full context in one fetch. His Organized AI hub is the master index. arch.organizedai.vip/claudeflare is the canonical example. I’m prototyping one for Freebo within two weeks.
Then I wrote this, exact letters:
Cloudflare framework - WHY CLOUDFLARE?! His upsell at his event is to host his people’s tools. Holy crap that’s genius.
Pattern #2. Jordan teaches Cloudflare-backed AI agent workshops, and at the end attendees can take their tool home and run it themselves, or hand it back to him to keep running. Students leave with something that actually works; Jordan stays in the loop with the people he just taught. It’s a genuinely beautiful design — teach the thing, then keep the door open. I’m borrowing the shape for the tour-operator vertical.
Pattern #3:
CACHING tokens. If done correctly could save hella
That’s the direct quote. Jordan’s word was “hella.” Cloudflare AI Gateway + KV is the play.
Pattern #4 — the Mediterranean ribbon-tied takeaway from his slide deck I wrote in panic:
TRY TO VISUALIZE THE ENTIRITY OF MY PROJECT IN THIS WAY: http://arch.organizedai.vip/claudeflare/#%23architecture
I also wrote:
Talk to Jordan about CHECKOUT
I will. Stripe Agentic Commerce inside the OpenClaw Gateway is its own thing. I’m collecting the question stack now.
Then this, which is just funny:
He has an openclaw training thing next week I NEED TO GO TO THIS /pa find it find him on the meetup and find his event
Yes the /pa is in the middle of my notes. That’s how I take notes now. I leave instructions for my own personal assistant inside the same document I’m jotting in. Twenty minutes later from my couch, the PA had found it: Clawstin May 21, 6–8 PM, Austin, registration open. It’s on my calendar.
”openclaw”
That’s the whole note. Just the word. Floating.
For about thirty seconds at home I thought I’d misheard “OpenClaude” — like an open-source Claude fork. No. It’s OpenClaw. It is its own product. Jordan’s product. Clawstin is the Austin meetup for it. There’s a whole Luma calendar of OpenClaw meetups in eight cities. I had to be told this by my own computer.
This part of the recap is my favorite because it’s the moment the system earned its keep — when an AI walked back a wrong assumption I’d loaded into it, because the URLs in my own notes contradicted me, and rerouted the entire research path. That’s not magic. That’s the unflinching middle of doing this well.
”Recursive self-improvement”
Translation: Julian Ghadially was up next. Founder of CodeEvolver. Self-improvement engine for AI reliability. The talk was a wall, in the best way.
Three takeaways, again my exact letters from my phone:
Takeaway #1: evolutionary is good, IMPROVE THE IMPROVER. DETERMINISTIC RULES ENGINES CAN OUTPERFORM LLMS. Takeaway 2: LLMS can synthesize rules engines that beat LLMS Takeaway 3: memory is not one and done
Capitalization is from the live notes. I wrote it in caps because that’s how I felt at the time.
Julian’s call to the room — and I am quoting him from memory, not from notes — was “if you care about development, you need evals.” Sample success conversations. Sample every failure. Codify both. Anything else is vibes.
He cited two papers worth reading back-to-back:
- The 2025 Sakana AI Darwin Gödel Machine — a coding agent that rewrites its own code. SWE-bench 20% → 50% open-ended. Polyglot 14% → 31%.
- ALMA: Learning to Continually Learn via Meta-learning Agentic Memory Designs — a Meta Agent that searches over memory designs as executable code, and discovers retrieval and update mechanisms that beat human-engineered ones across four sequential decision-making domains.
CodeEvolver is Julian’s framework for inspecting failures across all the levers — prompts, context, pipelines, reasoning steps, modules, memory, custom traces. I have my own thoughts on this. Mostly: my PA system is shipping new behavior every week and has no eval harness. That ends in May.
There’s a /pa directive in my notes from this moment too:
/pa make a task for me to follow up with julianghadially
The Todoist task is real. P3, due tomorrow.
”Nate - was walking agent orchestrator / subagents”
Nate Little, cofounder of SchoolAdmin (K-12 enrollment SaaS, multi-tenant, ships to thousands of schools). Walking the room with a demo of an agent orchestrator that splits subagent specialization by tech-stack layer — DB specialist, ledger specialist, API specialist, each holding its own context and tools.
That’s the design pattern I need for Freebo. Booking-flow agent should be one orchestrator with subagents on inventory, payments, comms. I’ve been hand-waving this for months. Nate has a working version.
I added him on LinkedIn before I left, then DM’d him the CopilotKit/skills repo because the architecture parallels are real. We’re getting coffee.
What my PA did with my AIMUG notes
OK technical reader. Here’s the part the AIMUG room will appreciate.
The whole stack: Claude Code (Opus 4.7, 1M-context) running locally as my primary terminal on Omarchy/Linux, plus an Obsidian vault, plus MCP connectors to Todoist, Google Calendar, Gmail, and a real Chrome browser. On top of that I’ve layered a set of custom skills — /pa, /pagoodmorning, /patasks, /paflow, /dbf — that encode the patterns I use over and over.
I dumped the raw AIMUG notes into Claude Code with /pa. Inside one session, it:
- Resolved every cryptic fragment. “Collin cases” → Colin McNamara. “openclaw” → not a typo of OpenClaude, actually OpenClaw. “Nate” → Nate Little, SchoolAdmin. It read the URLs in my notes, hit each presenter’s site and LinkedIn, and cross-checked against my existing vault.
- Wrote a 350-line Obsidian event log with TL;DR action list, a row per person, full speaker notes for Jordan and Julian, an open-source AI security tools roundup (Vigil, PentAGI, Decepticon, Agentic SOC — three of which I hadn’t heard of), and my verbatim gut reactions.
- Created a
People/*.mdfile per person. Frontmatter (type: person,met: 2026-05-06,linkedin_url: ...), a “why this matters” section, an engagement plan, a conversation log, and open questions. Colin’s note auto-taggedpriority: topafter the PA learned he founded the meetup. - Built a Networking Tracker that surfaces in my morning briefing (
/pagoodmorning) so warm contacts don’t go cold. The skill walksPeople/*.mdat 5 AM and surfaces anyone met in the last 14 days, the open Todoist follow-up tasks tied to them, and the next step pulled from each person’s engagement plan. - Created the right Todoist tasks in the right projects. DM Jordan (P2, due tomorrow). DM Julian (P3, tomorrow). Schedule coffee with Nate (P2, pinned, waiting). Audit my LinkedIn handle (P2, completed inline —
brettridenouris fine, no rebrand needed). Build the first PA eval set (P3, deep, 2h). Apply Jordan’s architecture-index pattern to Freebo (P2, Freebo / Product, 2h). Routing rules built into the PA put networking tasks inPersonal/Miscand Freebo deliverables inFreebo/Product. Inbox stays empty. - Calendared the right future events. Clawstin May 21. AIMUG June 3 (auto-detected the conflict with my recurring Wednesday Men’s Group and let me drop just that one occurrence). Vibe Code Austin May 20 (with a conflict noted in the description so I see it when the reminder fires).
- Drove a real Chrome browser for me. Verified my outbound DM to Nate by reading the actual LinkedIn message thread. Pulled my own profile to confirm my handle. Found Colin’s profile. Clicked More → Connect. Sent the request without a note. Confirmed the toast — Invitation sent to Colin — before reporting back.
End-to-end time, “here are my notes” → “everything captured, scheduled, ready for follow-up”: about 25 minutes. Most of that was the PA fetching pages over the network and me reading proposals back. It’s the same shape as the Search Console CLI I wrote earlier — a small tool with sharp edges, plus an agent that knows when to call it.
Without it: this all evaporates in 48 hours.
What I’m doing because of this one night
- Apply Jordan’s “Index of Everything Deployed” pattern to Freebo. One self-contained HTML page that drops a new collaborator into full project context in a single fetch. On the calendar.
- Build a first eval set for the PA. Julian made it concrete. Sample success conversations. Sample every weekly drift (the time my Todoist Inbox bloated; the day a morning briefing missed a high-priority email). Codify what “correct” looks like for each. Keep the success fixtures as regression tests.
- Cut LinkedIn shorts from Jordan’s keynote the moment the AIMUG YouTube channel posts the recording. The PA is watching for it with a check-back due May 20 and the
yt-dlpcommand pre-staged in the task description. - Show up for the next three. Vibe Code May 20. Clawstin May 21. AIMUG June 3.
The hype roll
Genuinely, thank you.
🥇 Colin McNamara — for hosting. For founding AIMUG. For the safety-as-a-constraint conversation at the bar. For building Acquit.ai while also running this entire Austin AI community. Your argument lives rent-free in my head.
🥇 Jordan Hill — for the talk that made me reach for my keyboard mid-sip. The Cloudflare hosting upsell is one of those reframes that, once you see it, you can’t unsee. Save me a seat at Clawstin May 21. I’ll be the guy with the laptop already open.
🥇 Julian Ghadially — for the eval gospel. For naming the Darwin Gödel Machine paper out loud. Improve the improver is now stenciled on my pipeline. The eval harness ships this month.
🥇 Nate Little — for the orchestrator demo that crystallized something I’ve been hand-waving for months. Coffee soon. I sent you the CopilotKit/skills repo on the way home.
If you were at the AIMUG Austin LangChain mixer and your name isn’t above — DM me. Especially Soschaba (the German hand-tracking guy whose gesture-only interface I want to see again — I have your name spelled wrong in my notes and I’d like to fix it) and Steve at Gastown (your orchestrator framework — I had thin notes and I want better ones).
The next AIMUG Monthly Mixer & Showcase is Wednesday June 3, 2026, 6–8:30 PM at ACC RGC 3000 in Austin. Same room. I’ll be there. Already on the calendar.
AIMUG FAQ
What is AIMUG? AIMUG stands for Austin LangChain AI Middleware Users Group, a 1,700-member community for AI builders, researchers, and middleware engineers in Austin, Texas. It was founded by Colin McNamara and runs monthly mixers, lightning-talk showcases, and hands-on demos.
Where does AIMUG meet? Monthly mixers are at ACC RGC 3000 (1218 West Avenue, Austin, TX) with an after-party at The Tavern. Many events are hybrid with a Zoom track. The official site is aimug.org; the Meetup group handles RSVPs.
When is the next AIMUG meetup? The next AIMUG Monthly Mixer & Showcase is Wednesday, June 3, 2026, 6:00–8:30 PM CDT at ACC RGC 3000. Mixers run roughly the first Wednesday of every month.
Who founded AIMUG? Colin McNamara founded AIMUG. He’s also building Acquit.ai (AI-powered litigation intelligence) and Self-Improving Code, and is a 25-year infrastructure veteran (ex-Apple iCloud, ex-Oracle Sr. Principal). USMC veteran.
What is Clawstin? Clawstin is the Austin OpenClaw meetup, run by Jordan Hill of Organized AI. The next Clawstin event is Wednesday, May 21, 2026, 6–8 PM in Austin — RSVP at clawstin.com.
How do I join AIMUG? RSVP via the Austin LangChain Meetup group or visit aimug.org. Membership is free.
This post was written by me. It was researched, scaffolded, fact-checked, cross-linked into my Obsidian vault, scheduled into three calendar events, and turned into eight Todoist follow-ups by the PA system I’m building on top of Claude Code. The ratio of human writing to system glue is approximately 1:1 and I think that’s the future of how a small team gets compound returns out of one person with a terminal. If you want to know how it’s wired, start here.