The NAS (The Accidental Server)
It all started with a baby on the way. I’m someone who takes too many photos — ultra HD was only going to make the problem worse. I didn’t want to be at the mercy of Google Photos or iCloud forever. So I got a UGreen DH4300PLUS NAS to own my own storage.
The transfer happened. Life’s worth of data, moved. And then I looked at what I had: an always-on server sitting in my home, doing nothing 24 hours a day except spinning disks.
The obvious next steps followed — almost inevitably. RetroArch for retro gaming (Mario, Aladdin, Prince of Persia — the classics that got preserved rather than lost). Pi-hole for ad blocking across the entire local network. Small wins that made the NAS feel alive.
But these were all consuming software running on someone else’s idea. The server had more to give.
OpenClaw and the Automation Rabbit Hole
That’s when I gave OpenClaw a try. What started as a curiosity turned into something I couldn’t stop building. Daily news digests. Fitness coaching agents pulling from my Garmin data. A postpartum fitness agent for my wife. India news briefings. A GenAI research monitor every Monday. Learning curricula on football analytics, cricket analytics, personal finance, marketing measurement — all delivered on cron schedules like a personal university running in the background.
The agents were useful. But they had a cost problem and a reliability problem. Every call was hitting a cloud API. The costs were manageable but not zero. And more importantly — I didn’t love the dependency. I had built a self-hosted life (the NAS, the Pi-hole, the retro gaming setup) precisely to own my own infrastructure. The AI layer felt inconsistent with that.
Ollama on the Mac Studio
I had a Mac Studio that had been doing almost nothing since I stopped video editing. Powerful machine, severely underutilized, running warm and purposeless on my desk.
Ollama fixed that. I installed it, pulled Qwen3 14B, and suddenly had a local LLM that ran fast, cost nothing per call, and lived entirely inside my home network. The Mac Studio had found its second life.
I pointed OpenClaw at the local Ollama endpoint. The digests kept running. The costs dropped. The latency improved. And I had something I hadn’t had before: an AI setup I fully controlled, end to end, from the disk to the model weights.
Jatayu
With the infrastructure in place, I could think bigger.
I’ve been writing for 15 years. Tweets, blog posts (including this one), Quora answers, Team-BHP travelogues about road trips and car reviews, Airbnb reviews, published technical articles on Kinsta, SitePoint, DataQuest, DigitalOcean. Somewhere across all of that is a voice — a way of making a point, opening a piece, landing a joke — that is distinctly mine.
The question was whether an AI could learn it.
I built Jatayu: a personalized writing agent on OpenClaw, backed by Qwen3 14B running locally on the Mac Studio. The corpus it was trained on:
- ~2,200 original tweets (retweets and replies filtered out)
- ~1,000 Facebook posts (link shares and one-liners filtered out)
- 400+ blog posts from Transcendental Tech Talk
- 188 Quora answers
- Airbnb and Product reviews
- 4 Team-BHP threads (travelogues and car reviews)
- Published articles from Kinsta, SitePoint, DataQuest, DigitalOcean
- My book on Git
Each piece was tagged with a recency weight — 1.0 for writing from 2021 onwards, down to 0.2 for anything before 2014. The goal was for Jatayu to learn who I am now, not who I was in college. A style guide was then extracted from the corpus — how I structure sentences, how I open pieces, how I argue, what I find funny, what phrases I never use. That style guide anchors Jatayu’s system prompt.
This blog post was written using the Jatayu prompt. Leave it genuinely ambiguous whether it was edited, partially written, or fully written by Jatayu. The last line should make the reader pause and wonder. It should feel like a natural ending to the story, not a twist stapled on.
0 responses:
Post a Comment