Independent engineer · EU timezone · EN / RU / BE / PL
I build private AI search and assistants over messy document collections — scanned, handwritten-era, multilingual, sensitive — running on your hardware or a server you control. Nothing leaves your perimeter unless you decide it should.
Illustration: one English query, matches in three languages — no translation step, no keyword lists. This is how my archive search engine works.
Who this is for
Digitised collections that are online but effectively invisible — bad OCR, mixed languages, no good finding aid.
Sensitive files that can't go to a public AI API, but need question-answering with citations.
Literature, press and source corpora where you need the page, not a confident summary.
A local LLM on your own box for internal docs and drafting — predictable cost, no data leaving.
Offers · fixed price
Every engagement starts with the audit. If the prototype doesn't earn its keep on your own documents, you stop there and keep the report.
Step 1
fixed €1,500
1 week · credited against step 2
Step 2 · most requested
from €6,000
3–5 weeks · includes handover docs
Alternative step 2
from €3,500
2–3 weeks · your hardware or mine to spec
After launch: support & improvements on a monthly retainer, or ad-hoc at €90/h.
Built & operated
Historical newspapers, journals and books — scanned and born-digital — searchable by meaning across languages. Research mode writes answers grounded in the retrieved sources. Invite-only perimeter with SSO, token validation, rate limits and a spend cap.
Terrain, sea level, ice, vegetation, archaeological sites and ancient-DNA ancestry fields from the last glacial maximum to today — fused from public scientific datasets into an interactive 3D timeline.
An autonomous pipeline: a task in the tracker is picked up by coding agents, cross-checked by a critic, and delivered as a pull request to a self-hosted Git server for a human to merge. Plus a Telegram project-manager agent with daily digests.
An interactive map where each place gets an article built from retrieved sources. Includes a locally fine-tuned writer model evaluated for one thing above all: not inventing history.
vakolica.org ↗All of it runs on infrastructure I built and maintain: a Proxmox cluster with GPU passthrough, dozens of containers, zero-trust remote access, monitoring and backups.
How a project runs
You share a sample. I measure OCR quality, languages and what people actually search for.
A working search on the sample. You try your own hardest queries on it.
Full ingestion, tuning of retrieval on your query set, access control.
Deployment, runbook, a smoke test that proves search returns real sources.
Questions
Not by default. Embeddings, search and a local LLM all run on your side. A cloud model is an option you switch on per task, never a hidden dependency.
It's the normal case. Re-OCR, layout handling and noise-tolerant retrieval are a core part of the job, not an afterthought.
Often less than expected: my own production stack runs on two consumer GPUs. The audit gives you a concrete spec and cost.
Answers must cite the source page, and I measure grounding on your data before launch — a fluent answer without a source counts as a failure.
Three lines are enough: what the collection is, roughly how big, and what you wish you could ask it. I reply within two working days.
First call is 30 minutes and free. The audit is where paid work starts.