shuffle-on
Thank you The GPU upgrade happened — 29 Sep 2026 Support the lab

Zero cloud. One machine. Real electricity.

A research lab in a 44 m² apartment in Iași: an offline library, a local AI agent and a live threat-intelligence platform. No cloud bill — the cost moved to the power meter.

Help keep it running What support already did
88CPU threads
503 GiBMemory
32 GBGPU memory
7Local AI models
160Workloads running
817Public dossiers

Figures refresh from the live system every 30 minutes.

What support made possible

The ceiling moved.

For months this page asked for one thing: more GPU memory. On 29 September 2026 it arrived.

Before

8 GB

An NVIDIA RTX 4060 Ti. Big questions crawled on the CPU or timed out; whole-book questions failed.

Now

32 GB

An AMD Radeon AI PRO R9700. The analysis that used to be queued, shrunk or skipped simply happens.

What it unlocked, in two weeks

A model router

7 local models behind one door, each woken in about 2.5 seconds when needed.

Fast everyday answers

The workhorse model answers at about 67 tokens a second on a single request.

Reading scanned books

Layout-aware OCR turns page images into searchable text, tables included — 1,163 pages so far.

Study guides

Section-by-section guides for 7,830 books, now written locally — no per-token bill.

A self-checking agent

lsn-agent-v3 plans, works through tools and flags its own unsupported claims.

Images on demand

Image generation and editing run beside the language models.

What it costs

The hardware was bought once. The power is every month.

An 88-thread machine with half a terabyte of memory doesn't idle cheaply — and it can't be switched off.

A honeypot that sleeps records nothing. A pipeline that stops leaves gaps.

This isn't funding to start something. It's help to keep running something that already exists and already publishes.

What comes next

Modest, and honest.

No moonshots. The things that keep the lab useful and safe.

  1. 1
    The power billFirst and least glamorous. It arrives whether the research goes well or not.
  2. 2
    Storage and backupsA library of 72,304 works deserves a second copy in a second place.
  3. 3
    Full-book study guidesToday the guides cover the opening of long books. Covering whole books takes GPU hours — which is electricity.
  4. 4
    Capability per wattEvery upgrade is chosen for what it does per watt, not raw speed — otherwise it makes the bill worse.
What does not change. Nothing goes behind a paywall. The dossiers stay public, the methodology stays published, the disclaimer stays strict. No supporter tier, no early access, no private feed. Support buys nothing — it just means the work continues.
How to help

Any amount. No account needed.

Send a contribution

One-off or recurring. Neither link asks you to create an account.

☕ Buy me a coffee PayPal

Rather not send money?

  • Read the dossiers and tell me where the reasoning is weak — corrections are worth more than coffee.
  • Share them with someone who works in this field.
  • If a dossier describes infrastructure you operate, write to me: lisneanucristian@shuffle-on.com or lisneanucristian@gmail.com
Work together

Need something like this for your own team?

Private AI that runs on your hardware, an offline library, a teaching platform, a threat-intelligence pipeline — I build and run these. If any of it would help you, let's talk.

Talk to me on LinkedIn Read the disclaimer

Personal site. Research and opinions are my own.