May 2026
Date: May 2026
Topics: AI, job search, local LLMs, automation, career
---
When I started looking for my next role, I did what I've always done when faced with a messy, multi-dimensional problem: I built a framework.
For the last 25 years, that's meant building frameworks for other people — cloud migration patterns at Micro Focus, professional services operating models, SRE team structures. This time, it was for me.
Three months ago, I set out to automate as much of my job search as possible using local AI agents. No ChatGPT subscription. No API keys. Just open-source models running on hardware I already own.
Here's what the stack looked like, what worked, what hallucinated, and what I learned about applying the same enablement framework to my own career.
---
I run three machines in a home lab:
The key piece is something I call OB1 — a local database that stores everything as "thoughts" with vector embeddings. When I feed in a job description, it doesn't just match keywords. It searches across 100+ entries from my career notes, past blogs, and project documentation, then returns the most relevant experience to talk about.
No data leaves my network. No API calls. No privacy concerns.
[PLACEHOLDER — Insert specific moment when you realized this would work. Example: "The first time I fed a JD through and it surfaced a meeting note from 2021 I'd completely forgotten about — a call with Erez about Kubernetes scaling — I knew I was onto something."]
---
The most useful thing by far wasn't generating applications — it was analyzing job descriptions against my actual experience. I'd feed a JD into the pipeline and get back:
For the first time, I could triage 20 job postings in 30 minutes instead of 3 hours. The tool doesn't decide which ones to pursue — but it tells me which ones are worth a deeper look.
[PLACEHOLDER — Insert a concrete example. "For the Zscaler role, the AI flagged that my CloudFlare Tunnel work maps directly to Zero Trust architecture — something I wouldn't have thought to lead with."]
The interview answer bank I've been building — structured STAR responses tied to specific career events — lives in OB1 as vector embeddings. When I'm preparing for a specific company, I search across the knowledge base with the company name and get back every relevant project, meeting, and result.
The AI doesn't write my answers. It finds the evidence I've already documented. I still have to connect the dots and deliver it in my own voice. But I'm no longer staring at a blank page trying to remember what I did in 2018.
This blog exists because of the same pipeline. Every post is reconstructed from 12+ months of private OneNote journal entries and a 45-file archive of corporate strategy materials. The AI finds the narrative threads — I rewrite them to sound like a human wrote them.
The result: 8 posts in 14 days, all grounded in actual events, not generated from thin air.
---
This was the biggest problem. More than once, the AI generated confident-sounding bullets about projects that never happened, certifications I don't hold, and results I can't prove.
Example: It wrote that I led a "global Kubernetes rollout across 47 accounts" — which sounds plausible until you realize the actual project was about AWS Landing Zones, not Kubernetes. The AI filled in the technical details from its training data, not from my notes.
The fix: never use AI output without verifying against primary sources. My OneNote journal is the source of truth. Everything gets checked.
I tried fully automated applications — the AI writes the cover letter, fills the form, hits submit. The results were bad. Generic, obviously templated, no signal of genuine interest.
The applications that got responses were the ones I wrote myself, the ones where I named specific people, referenced specific conversations, and meant what I said. The AI can draft. It can't care.
It's tempting to set up a pipeline and let it run — scrape LinkedIn, analyze postings, generate outreach. But job search is a relationship business, not a throughput game. The automated approach generated volume. The human approach generated interviews.
[PLACEHOLDER — Insert a short anecdote about a real human interaction that got results. A referral, a coffee chat, a message that worked.]
---
None of this was new. I was applying the same model I used at Micro Focus:
1. Assess — Know where you stand (fit scores, gap analysis)
2. Standardize — Build the patterns (answer bank, cover letter templates, prep workflow)
3. Enable — Hand over the playbook (in this case, to myself)
The technology doesn't change the fundamentals. The CCoE framework for enabling product groups to adopt cloud is the same framework I used to enable myself to run an effective job search. The domain is different. The pattern is identical.
[PLACEHOLDER — "The first time I realized this, I laughed out loud. I'd spent four years building frameworks for other people to scale, then spent three months building one for myself without noticing the irony."]
---
Two weeks ago I started publishing these posts. Traffic so far: 272 page views from 10+ countries, with a sustained base of 22-26 unique visitors per day. Not huge — but it's growing.
The bet I'm making is that local AI infrastructure isn't just a job search hack. It's a model for how enterprise AI operations should work. Private. Controllable. No per-query costs. Built on open-source models that run on hardware you already own.
The CCoE playbook applies here too. First you build the reference pattern. Then you enable the teams. Then you scale.
Right now I'm on step one.
---
*Next up: How I built the OB1 memory system in an afternoon — the actual architecture, the SQLite schema, and why every engineer should own their own knowledge base.*
---
About the author: Andrew Katana spent 25 years building the frameworks that let engineering teams scale — from reference architectures at Novell, to PS operations at Micro Focus, to the Cloud Center of Excellence that took SaaS revenue from $240M to $450M. He's currently applying the same model to AI infrastructure. Find him on [LinkedIn](https://linkedin.com/in/andrew-katana-615294).
Built on a home lab, powered by local models, and owned by Andrew Katana.