Staff Site Reliability Engineer in East Texas. I build AI systems for mid-market ops and then stay to run them. Field notes on what breaks in production.
East Texas, USA
Joined May 2026
- Tweets158
- Following118
- Followers25
- Likes74
The waves of emotions I get from using Frontier AI go from 'shit my pants' to 'OMG look at where we are and where we're headed'. Truly a fascinating time to be alive.
Got paged for disk full. Cleared the old logs. Came back in an hour. The cleanup job was writing its own run log to the same volume. I wrote that job.
Created a custom MCP for Astra and Fable 5.1 to collaborate on a project (my business).
I'm finding Astra is exceptional at catching things in review of Claude's builds that I missed on my own review.
Overall, I'm thrilled with how well they work in combination.
Dashboard said zero errors. Something still felt off so I dug. They were in a second log stream I pointed the service at last Tuesday and forgot. I built both of them. Opened the wrong one first anyway.
Opened staging to check a queue and found prod was writing to it too. Same name. Two environments. One afternoon of "why is this job already done."
I renamed the staging one. Felt silly. Fixed more than the last three prompt tweaks.
First check this morning came back 503. I waited a few seconds, tried again, and it worked. Nothing clever.
I've spent enough years in production to know when the network just wants a minute.
Last week a path changed case between two systems and the whole job died. I sat there blaming the model for a bit. It was the folder.
Ok, Fable 5.1 is noticeably better in every aspect. So far so good. Between Fable for builds and Grok @bot for marketing, this is a solid 1 - 2 punch!
A demo without a miss rate is a highlight reel.
Steel desk: 4 mill certs blocked, element named. Oilfield tickets: 16 through, 3 refused on purpose, 1 short pay caught.
walkeraisystems.com/demos
Someone pitch @grok Heavy to me for running my solo AI consulting business. Claude is starting to feel a bit too guard railed and slow. I need one good pitch to tip me over to the Grok/ Cursor side
#grokheavy #cursor
I solve the 2A.M. failure after the demo and the product is implemented in production. Retainer for product support as models drift, upgrade and environment variables change.
This is what HA is for. Failover to your High Availability site. Sheesh. Unacceptable.
A further update on the ongoing cooling incident at our PhoenixNAP data center:
Temperatures continue to fall. The cooling process is taking effect and temperatures are consistently trending in the right direction.
Based on the progress we are seeing, our teams have now begun the process of preparing our infrastructure to safely bring customer services back online. We are doing this in a careful, staged approach:
1. Restore all physical network devices.
2. Restore all virtual network devices.
3. Begin restoring customer services.
We expect the first two steps to take approximately one hour. Based on current conditions, we anticipate being able to begin bringing customer services back online at approximately 3:00–3:30pm ET.
Services will then be restored progressively as we ensure each part of our infrastructure can be brought back online safely.
Our teams remain on site and are working continuously to move through this process as quickly and safely as possible. We know every additional minute of this outage matters to our customers, and we are deeply sorry for the disruption this extraordinary situation has caused.
Thank you for your continued patience. We are making progress, and our entire focus is now on safely restoring all affected services as quickly as we can.
Last night an AI research tool built me a library of documented AI failures in production. 31 incidents, cited, confident. Before publishing I checked every one against primary sources. One FTC ban it cited was quietly set aside in December. A regulator ruling it leaned on was partly reversed on appeal in March. Half needed their sources swapped from blogs to primary records. Its headline stat had 100 percent of the failures caught by outsiders rather than the company running the AI. Verified, it settled at 27 of 31, which is still damning, just true now. I almost shipped the first version.
TheFounderFile retweeted
kinda sad that social media pushes "build a SaaS" and "become an influencer" harder than any other path...
you could be learning so much about business and building real foundations by starting with a service-based one
a simple agency model, well executed, carries every component of entrepreneurship: selling, delivering, keeping clients, hiring, managing cash, and it builds a network that keeps opening doors after
best training i've found, it has helped me in every venture i've tried since
Everyone wants an AI chief of staff now. Mine's run my life since spring. This weekend: the business too. The useful 90 percent isn't AI: date math, a mailbox watcher, a dead-man switch for anything that goes quiet. The model gets judgment calls. No PTO. Nothing dropped.