Selected work/

A multi-user agent setup, running since the early days of OpenClaw

I started running OpenClaw close to its launch, when agentic assistants were still new, and have operated it daily since. It now spans two servers, my own VPS and a GPU workstation at work, with several people, agents and channels sharing one carefully scoped setup.

Searching 5,150 technical documents in two languages

Product facts lived in data sheets, safety sheets and manuals, not on the product pages. Only 5 of 18,673 product records linked a PDF. I chose to build search first and leave document generation for later.

118,387

indexed page-chunks

~58,700

pages, OCR only where needed

8 / 8

typo queries found first

0 / 81

broken citation links (chosen model)

How it works. Search combines dense embeddings with BM25 and merges the two rankings with reciprocal rank fusion. Accent-free forms are indexed, so Icelandic searches work without perfect accents. Queries are also translated to English, but the original query is always searched too, because the translator sometimes misread Icelandic. The assistant searches first and cites document, page and source PDF.

Decisions made on evidence. A reranker was tested and switched off because it made Icelandic-to-English retrieval worse. In 24 paired answers, the slower model (about 10.6 s vs 3.8 s) searched and cited every time with no broken links. The faster one had 9 broken links out of 49. I picked the reliable model. A user-editable glossary fixes Norwegian and Danish words that leaked into Icelandic answers.

Limits: On a 38-question set the right document ranked first about 29–30 times. That measures retrieval, not answer accuracy. The model comparison was a preference test, not a blind trial.

A local AI server built around 32 GB of VRAM

I wrote the build plan for an RTX 5090 workstation. The goal was a working environment, not a model demo: local LLMs, document knowledge, creative generation, internal apps, website staging and agent orchestration.

The design starts from the real constraint, which is GPU memory. One large chat model and one embedder stay resident, and creative jobs are queued and unload the chat model when they need the GPU. The system uses retrieval rather than fine-tuning. Mismatched context sizes forced models to reload, so we aligned them. Hot data is kept on NVMe and archives on a separate SMR disk.

Diagnosis is organised in four layers: private network and TLS, Windows-native processes, WSL/Ubuntu services, and two separate Docker engines. Something can work in its container and still fail from Windows, or appear missing because the wrong engine was queried. Startup sequencing and a WSL keep-alive were tested with cold starts.

Ollama · Open WebUI · Qdrant · PostgreSQL · Redis · Docling · n8n · Caddy · OpenClaw agents · Tailscale

A creative studio that knows when not to generate

The studio puts local image and video generation behind a product interface (React, FastAPI and ComfyUI), so nobody has to edit node graphs.

FLUX.2 Klein handles image creation and editing. LTX handles video: image-to-video, restyle, continue a shot, follow motion, and grab a frame as a reference. Seeds and settings can be locked, so results can be reproduced.

The product-image page is deliberately non-generative. Distributor photos are framed on consistent square canvases with a set object footprint, normalised white backgrounds, and Real-ESRGAN upscaling only for small sources. A generative model can redraw packaging text, so it isn’t used where the product has to be shown exactly.

Boundaries: Rendering and source media stay local, but assistant messages go to a cloud provider. One video model was archived without being used because its license excluded our region.

AI tools for two shops, with a human approving production changes

Assistants get typed MCP tools for the catalogue, content, taxonomy, analytics and webmaster work. They do not get general access to the stores.

• One store server per shop. The only service that can read both shops is read-only by construction.
• Page tools write drafts. Editing a live page creates a linked draft for review, and product-feed changes go through an approval page.
• OpenSEO was adopted rather than rebuilt. Agents use its 52-tool API alongside Search Console and GA4.
• Bing Webmaster has 16 tools, and the only writes it allows are recrawl submissions. In its first review we found one URL that was really missing and redirected it, then submitted 48 URLs.

Code is built on dev mirrors and checked with Playwright on desktop and narrow phones before anyone approves it for live.

From asset library to marketing workbench

An in-house DAM (Node and Express) took on uploads, previews, folders, ad sizes, a booking calendar, delivery reminders and expiring partner links.

8.0 → 2.7 s

per-PDF thumbnail render

228 ✓ / 10 skip

regression suite on the UI refresh

Large upload queues broke on a per-request file cap. The client now splits the queue into server-sized batches and shows one combined progress bar. A separate state shows server-side processing after the upload finishes. Concurrent writes were reworked to read fresh state before committing. Filenames keep Icelandic characters on display and are transliterated for URLs.

Editable WordPress without visual regressions

A B2B parts catalogue (industrial supplier) has 3,049 products and 468 categories in development. A hand-built front page was converted into custom server-rendered blocks, with zero changed pixels at seven viewport widths. Search runs on a site-specific Meilisearch integration with an accessible combobox. Ties in ordering were made deterministic, so visual tests stay stable.

A local auto-service website went from my brief and logo to a live site. The front page is built from real editable blocks rather than freeform markup. It was tested at desktop, tablet and 390 px. The mobile call button was then designed down to 320 px. Afterwards we closed the project out with an archive, a git bundle and restore notes.

A traffic spike that wasn’t growth

After a tagging and consent change, GA4 showed a sharp rise in traffic. We compared it with Search Console and commerce data and called it what it was: a change in measurement, not more visitors. The same data later fed a homepage brief covering mobile share, search behaviour and top categories.