Bare Metal and Neural Nets: Crafting a Purpose-Driven AI Cyberdeck

A developer built a purpose-driven AI cyberdeck that runs large language models locally without cloud dependencies. The build uses a Raspberry Pi 5, NVMe storage, a Hailo-8 AI accelerator, and a Raspberry Pi Pico W for power management, enabling offline inference, transcription, and RAG. The project prioritizes cognitive sovereignty over aesthetics, rejecting the 'cosplay' trend of non-functional cyberdecks.

Your laptop is a rental. Your cloud is someone else's computer. It's time to build something that is actually yours. Everyone has seen one. A mechanical keyboard crammed into a Pelican case, a tiny screen running cmatrix for the 400th time, a cyberdeck that has never left a desk. It's beautiful. It's also useless. I love the aesthetic as much as the next Gibson-pilled gremlin, but we have to be honest: 90% of cyberdecks are cosplay. They are props for a future that never arrived. They look like they could decrypt a corporate mainframe while the city burns outside. In reality, they struggle to open a PDF. I got tired of building toys. I wanted to build an exocortex. Something I could throw in a bag, take off-grid, and actually think with. No API keys, no subscription, no "Sorry, our servers are experiencing high demand." Just bare metal and neural nets, solder and inference, purpose over performance art. This is how you build one that matters. The cyberdeck community has fallen into what I call the Blade Runner trap. We optimize for how a thing looks in a photo on Reddit, not how it performs when the Wi-Fi dies. The classic build: Raspberry Pi 4, underpowered, thermal-throttled inside a 3D-printed box that took 48 hours to print and 10 minutes to become annoying. It's a vibe. Vibes don't run 7B parameter models. A purpose-driven deck starts with a different question. Not "What looks cyberpunk?" but "What problem do I want to solve when no one is watching?" For me, that problem is cognitive sovereignty. I wanted a machine that could do three things without ever phoning home: If your deck can't do those three things, it's a keyboard in a lunchbox. No offense to lunchboxes. This is where we stop building props and start building tools. Purpose is a protocol. Everything else is decoration. Let's talk guts. If you want local AI that doesn't feel like you're running ChatGPT on a toaster, you need to be ruthless about your bill of materials. The Brain: Raspberry Pi 5. Yes, the Pi 4 was the darling. The Pi 5 is the divorce. It's actually fast. PCIe exposed. Real I/O. It doesn't beg for mercy when you ask it to compile anything. For a cyberdeck, this is your baseline, not your stretch goal. The Spine: NVMe. If you are still booting your cyberdeck from an SD card in 2026, you are in an abusive relationship with storage. SD cards are slow, fragile, and corrupt at the exact moment you need them. An NVMe base I use the Pimoroni NVMe Base, but there are a half-dozen good ones now changes everything. Boot times go from geological to instant. Vector databases stop crying. You can actually run an OS that expects to be treated with respect. This one upgrade turns a toy into a workstation. The Occult Co-processor: Hailo-8. This is the secret weapon. The Pi 5's CPU is good, its GPU is okay, but neither was built to do 26 TOPS of AI inference without melting. The Hailo-8 AI accelerator is a little M.2 module that sits on that PCIe lane you just freed up and gives your deck a dedicated neural engine. We are talking real-time object detection, zero-latency transcription, local RAG that doesn't take 45 seconds to answer "what did this PDF say?" It sips power while it does it. This is how you get offline AI that feels online. The Nervous System: Raspberry Pi Pico W. Why add a $6 microcontroller to a $100 computer? Because smart systems have dumb reflexes. The Pico W is my power manager, my battery brain, my graceful-shutdown handler, my hardware kill-switch for the radio stack. It watches voltage, it controls fans based on actual thermal curves, it lets me wake the main Pi with a physical key. It means my cyberdeck doesn't die like a laptop. It sleeps and wakes like an animal. Also, having a second, fully air-gapped brain that can run Meshtastic or act as a BadUSB-detector in the field is just fun. Redundancy is punk. Power is politics. I run a 65W PD battery bank with a proper UPS HAT in front of it. The deck can hot-swap power without blinking. Thermals are handled with a low-profile blower and actual airflow, not just holes drilled as an afterthought. The case? Pelican 1150, foam ripped out, custom deck plate cut from 3mm carbon fiber. No 3D printed hinges. If it can't survive being checked on a bus, it's not portable. Hardware is just expensive paperweight without the right software. The goal is a local-first, offline-capable cognition machine. No cloud. No telemetry. No Terms of Service that change next week. My OS of choice is a stripped-down Raspberry Pi OS Lite 64-bit, no desktop bloat. I run Sway as a tiling window manager because it weighs nothing and looks like it belongs in The Matrix without trying. On top of that, the real stack: 1. The Model Layer: Ollama with Llama 3.1 8B quantized Q5 K M as the daily driver, and Phi-3 Mini for fast tasks. With the Hailo-8 handling embeddings and vision, the Pi 5 CPU is finally free to just... think. No more watching tokens trickle out one by one like a dying IV drip. 2. The Memory Layer: This is where most people fail. An LLM without memory is just a very confident amnesiac. I run Qdrant as a local vector store, fed by everything I research. PDFs, HTML captures, field notes, transcripts. All chunked, embedded locally with nomic-embed-text , and searchable offline. The LLM doesn't hallucinate when you give it receipts. 3. The Interface Layer: Open WebUI but heavily modded. Or even better, Obsidian as the front-end. Every chat, every research session, every transcription is a Markdown note. Linked. Tagged. Graph-viewed. Your deck becomes a second brain that actually has a brain stem. This is the part where people usually get lost in tooling and forget why they started. I did too, for months. I had a beautiful pipeline that could do everything except save me time. The breakthrough was realizing that AI is not the product. The workflow is the product. For investigative work, for example, the difference between a cool demo and a useful tool is having your pivot logic already mapped. I got tired of re-building the same OSINT chains every time I went down a rabbit hole, so I finally codified all of it — every username pivot, every dork that Google pretends doesn't work anymore, every Obsidian template I use to turn chaos into a case file — into what I now just load onto every deck I build: the OSINT Investigator Pack: Pivots + Dork Bible + Obsidian Workflows + AI Toolkit https://numbpilled.gumroad.com/l/osint-investigator-pack . It's not sexy. It just means when I'm offline in a cafe with my deck, I can go from a handle to a network map without touching the cloud. That's what purpose-driven means. The AI doesn't just chat. It executes a workflow you actually use. Let's be brutally practical. A purpose-driven build has stages. Don't try to do it all at once or you'll end up with a drawer of half-soldered HATs. Stage 1: Prove the Brain. Pi 5 + NVMe + display + power. Get it booting from NVMe reliably. Get your OS hardened. If you can't make this stable, the rest is theater. This alone takes an afternoon because Pi bootloader docs were written by sadists. Stage 2: Add the Muscle. Install the Hailo-8. Get the HailoRT drivers playing nice with your kernel. Run your first local inference benchmark. Feel the moment when your object detection goes from 2 FPS to 30 FPS. That feeling is what the internet stole from us. Stage 3: Give it a Soul. Install your local LLM stack. Build your first local RAG pipeline with three PDFs you actually care about. Not Wikipedia. Your PDFs. Your notes. If it can't answer a question about your work, it's not your AI. Stage 4: Make it an Organism. Integrate the Pico W. Program the power logic. Add the physical switches. The moment you can kill Wi-Fi/Bluetooth with a physical toggle and your deck keeps thinking, you'll understand why cyberdecks matter. Privacy is not a software setting. It's a hardware switch. I get DMs daily asking for the exact wiring diagram, the config.txt that doesn't brick the NVMe, the Sway config that makes a 7-inch screen usable, and how to cluster two Pi 5s inside one case without creating a space heater. I was going to write a 10-part thread, then realized no one wants a thread, they want a build log that actually boots. So I dumped the whole thing — full BOM with current prices, STLs that don't suck, the Hailo-8 + Pi 5 NVMe dance, the Pico W power-brain firmware, and the cluster setup I use when I need one deck to be a router and one to be a brain — into the Offline AI Cyberdeck Pack: Pi 5 NVMe + Hailo-8 + Pi 5 Build + Pico W + Cluster https://numbpilled.gumroad.com/l/offline-ai-cyberdeck-pack . It's the build guide I wish I had six months ago. Here's the irony: when you build for purpose, it ends up looking more cyberpunk than the cosplay decks ever could. My deck is scuffed. The carbon fiber has a burn mark where I shorted a buck converter. The keyboard is a beat-up Corne split that I never learned to touch-type on properly. There are stickers covering mistakes. There is a single, unnecessary, extremely bright red LED that tells me when the Hailo-8 is inferring. It serves no functional purpose. I love it. True cyberdeck aesthetics aren't about looking like a movie. They're about evidence of use. Patina. Wear patterns on the keys you actually press. A handle that is exactly where your hand wants it. The beauty of a tool that has been used to do real work is fundamentally different from the beauty of a prop. McLuhan said we become what we behold. We shape our tools and thereafter our tools shape us. If you build a tool that requires you to be tethered to Big Tech's teat to function, you become a dependent. If you build a tool that is sovereign, you start to think like a sovereign individual. Your cyberdeck should be a little hostile. Not to you, but to the idea that you need permission to compute. So stop optimizing for Reddit karma. Stop printing another hexagonal grill. Define a job that matters to you. OSINT researcher who needs to work from a parking lot. Writer who wants an AI editor that never reads their draft. Field technician who needs vision models in a basement with no signal. Ham radio operator who wants to transcribe and summarize signals offline. Pick one. Build for that one person. Bare metal doesn't lie. Neural nets don't need Wi-Fi. The future isn't going to be handed to you in a polished aluminum slab from Cupertino. You're going to have to build it yourself, in a Pelican case, with a soldering iron that is slightly too hot, and firmware that almost works. And when it finally does boot, when that local model answers a question about your own notes without ever touching the internet, you'll feel it. That's not nostalgia. That's power. Build your exocortex.