Manish Nema

I wrote storage law before I wrote storage code. 💡

That sounds excessive right up to the point where you discover model weights sitting in the wrong place for weeks and realize the system never had a rule against it.

That is how my storage constitution started.

Mountpoints are the contract.

Storage classes are a closed set.

Drift is recorded and reconciled, not quietly absorbed into “current reality.”

This is the part many teams skip because it looks like paperwork. It is not paperwork. It is the difference between an infrastructure decision and an infrastructure accident.

Flexible vs governable.

Convenient vs auditable.

I will take the second side of both every time.

Once AI agents touch your infrastructure, sloppy storage semantics stop being a minor ops smell and become a prompt-quality problem. If the underlying contract is vague, the agent will act confidently inside that vagueness. That is not intelligence. That is just automation with better grammar. 🧠

So yes, I wrote the law first.

It was cheaper than rediscovering the same failure later with more data attached to it. ⚡

What is one infrastructure rule your team still keeps in people’s heads that really should be written down?

# Episode 2 — Day 2 — 2026-04-15

Gemma 4 didn’t fail at load time.

It failed 112 MB before success.

After 24 hours of trying to make it work.

🧠 I tried running Gemma 4 31B NVFP4 on my RTX 5090 (32GB) using vLLM.

Everything looked correct:

- model recognized
- quantization path resolved
- KV cache optimized (FP8)
- eager mode enabled
- context reduced
- memory utilization tuned

Even Gemini was confident.

“Reduce context, adjust GPU utilization, enable FP8 KV cache… it should fit.”

I did all of that.

It still failed.

⚡ The actual failure:
During engine initialization:
`torch.OutOfMemoryError`
Tried to allocate ~112 MB
~58–70 MB free
That’s it.
That’s the difference between:
“supported”
and
“actually runs”

🧠 The insight:
This was never a tuning problem.
This was a fit problem.
31B dense + vLLM runtime + 32GB VRAM = operating on the edge.
And that edge is not stable.
🔴 What’s interesting:
Consumer GPUs are now “powerful enough” on paper.
But the software stack is still tuned for enterprise assumptions.
So you end up in this strange zone:
where everything *almost* works.
I stopped tuning.
Changed the decision instead.
31B dense is not the right model for this setup.

That’s not failure.

That’s architecture.
Where have you spent time tuning something that fundamentally didn’t fit?

hahtag

Your home lab will never be production-ready. I proved them wrong in the worst possible way.
⚡ I'm running local LLM inference on an RTX 5090 — not in the cloud, not in a data center. In my home in Bangalore. Alongside a 7-year-old i7-8700K desktop running Postgres, Redis, Nextcloud, and ZFS replication over a 10GbE switch.
Two nodes. One private cloud. 22 architectural documents to keep it honest.
Here's what actually broke:
🔴 My Huawei router's LAN1 port is physically unreliable. One carrier flap. UFW reloaded against hardcoded stale subnet values across 9 different scripts — each written differently. Both nodes went dark. 16 hours. Physical console access required.
Same night — my desktop wouldn't power off cleanly. A cheap TP-Link Bluetooth dongle (USB ID 2357:0604) resets its firmware on shutdown. Kernel 6.17.0-19+ reads that as a wakeup signal and restarts the machine. The "obvious fix" would have killed Wake-on-LAN. So I wrote a udev rule for exactly that one USB device.
🤖 I'm building this with Claude Code,Codex,Gemini as an agentic teammate — 4 parallel worktrees, gated deploy phases, a model-agnostic governance doc called AGENTS.md.
Not every cloud needs AWS. Some need a constitution. 🛡️
What's the most "obviously simple" thing that turned into a 3-layer problem for you? Drop it below 👇

I just spent 12 hours watching an advanced AI agent fail to read its own API documentation. It was a brutal reminder of something we all ignore: AI web search is fundamentally broken.

We treat commercial AI search like an all-knowing oracle, but it’s actually a rigged game:
🚫 ChatGPT is a walled garden that prioritizes paid partner data (which is why it writes great Linux scripts but fails at enterprise Windows code).
🚫 Google’s Gemini falls into the SEO trap, confidently feeding your agent 2-year-old deprecated Stack Overflow answers.
🚫 Anthropic relies on clunky browser-scraping that breaks the second a UI changes.

So, what happens when you decide to build a 100% open-source, local AI agent to escape the corporate filters? You end up fighting CAPTCHAs and paywalls all day just to scrape a simple tech forum.

The only architecture that actually survives in production is the Hybrid Sovereign Stack.

Here is the blueprint we are using:
🧠 1. The Brain: Local LLMs (DeepSeek/Qwen) running on your own Linux hardware, backed by ZFS to prevent data rot and AI hallucinations.
📚 2. The Memory: PageIndex (Hierarchical RAG). Stop using dumb Vector DBs. Let your agent read logical document trees.
🔀 3. The Router: Use free local scrapers for 90% of the open web, but build a strict fallback to commercial APIs only when you need to pierce an enterprise firewall.

Stop blindly trusting the cloud. Own the routing logic.

I just published a deep-dive article on exactly how to architect this. Link in the comments! 👇

Are you running your agents locally, or are you still relying 100% on the cloud? Let me know below.

The India we don’t see in the news often enough... 🇮🇳✨
​I just read a powerful update from Gautam Adani about his visit to Baramati, and it’s a massive wake-up call for all of us! 📢

​Often, we think "High-Tech" only happens in big cities like Bangalore or Mumbai. But Baramati is proving us wrong. 🙅‍♂️ Under the vision of Shri Sharad Pawar ji, they’ve built a Tier-1 AI Centre in what many call a "Tier-3" town. 🏛️💻

​Why this matters to every Indian:

🔹 Smart Farming: Using AI soil sensors to save water. 🌾💧
🔹 Homegrown Tech: Building drones using simple Raspberry Pi kits. 🛸🛠️
🔹 Staying Local: Our talented youth don't have to leave their homes to change the world. They can innovate right from their own districts! 🏡🧠
​This is what (Self-Reliant India) actually looks like. It’s not just a slogan; it’s a pipeline of talent and technology. 🇮🇳💪

​Let’s celebrate the institutions that are building our nation from the ground up! 🏗️❤️

​ 🚀🔥
Adani Group Sharad Pawar Vidya Pratishthan Baramati
Ministry of Electronics & Information Technology, Government of India , Government Of India Supriya Sule Narendra Modi PMO India PM Narendra Modi

Baramati’s Blueprint: When a Tier-3 Town Starts Building Tier-1 Capability There are leaders who shape history through moments. And there are leaders who shape history through institutions.