case study

The assistant living on this site is not an API wrapper. It is my own fine-tuned model, self-hosted on a modest ARM server, answering with live retrieval — and it exists precisely to prove that this class of system is buildable and affordable.

  • Model Gemma 4, fine-tuned with QLoRA
  • Hardware 4-core ARM server, no GPU
  • First response 30–45s full answer, words from ~20s

  1. The problem

    Every "private AI assistant" pitch runs into the same buyer suspicion: does this actually work without OpenAI behind the curtain, and what does it cost to run for real? The only honest answer is a running system a stranger can poke. So the studio's own assistant became the case study: fine-tuned, self-hosted, public — right here on this page.

  2. The constraint

    One small rented server, no accelerator — and every constraint a real client brings, stacked on top deliberately: a small open model (Gemma-class, not a frontier API), bilingual EN/RU output, answers grounded in the studio's actual facts rather than the model's imagination, and graceful degradation, because the site has to work in full when the AI box is switched off.

  3. What was built

    A Gemma 4 fine-tuned with QLoRA on a purpose-written dataset — behavior and voice come from the fine-tune, facts come from retrieval (RAG) over a curated knowledge base: tune for form, retrieve for truth. It serves through llama.cpp behind a tunnel, with Google sign-in and a message allowance guarding the expensive path. Chat history lives twice — IndexedDB in the browser, SQLite on the server — kept in sync so a conversation survives closing the tab. And when the box itself is unreachable, the frontend quietly degrades to a scripted mock instead of a broken widget: autonomy by design, not an afterthought.

  4. What it cost

    Time, not a number on an invoice. Months of iteration went into a purpose-written training dataset that grew through several generations before the model answering you now was allowed to ship, evaluation gates every candidate had to clear first, and at least one full retrain that failed its own gate and was shelved rather than deployed. Nothing here talks to a real visitor until it passes the same kind of acceptance test a client's own pilot gets — that discipline is the actual cost, and it never shows up as a line item.

  5. What it does today

    On four ARM cores a full first answer takes 30–45 seconds; first words land around the twenty-second mark. Later turns in the same conversation are faster, because the context is already read. That number is published, not hidden — it is the honest price of CPU-only sovereignty at the smallest possible budget. A client deployment sizes hardware to the latency the workflow actually needs; the same architecture on a modest GPU answers in seconds.

  6. What it doesn't do well

    Three honest limits. It is slow by design — a CPU-bound small model will never race a frontier API, and you feel that on the very first turn. It is narrow, not general — ask it something outside the studio's own knowledge base and retrieval has nothing to hand it, so it says so instead of guessing. And voice is one-way for now: it can speak a reply back with Google's TTS, but transcribing your voice is the piece still unwired — type instead of talking, for the moment.

What this proves for your project

That a fine-tuned small model plus retrieval delivers a domain assistant without per-token bills or data leaving your infrastructure, running on one rented server rather than a GPU cluster — and that the person offering to build yours has already built one in production, in public, with the trade-offs published instead of hidden. Click the brain on the homepage and interrogate it yourself. If your documents need the same treatment, start with the RAG pilot — a fixed three-week engagement; if compliance means the model has to live on hardware you own outright, that is self-hosted AI . Both sit on the full price ladder .

Ready when you are

Write two sentences about what you're trying to do. You'll get a straight answer — including "you don't need me for this" when that's the truth.

Start a conversation

Contract & NDA before any work beyond the audit · milestone payments, typically 30/40/30 · 30-day warranty · registered business, Armenia · GMT+4, EU hours · Trust & Process