Selected work/

Shared agents for everyday work

A shared OpenClaw setup gives colleagues access to assistants for website work, imagery and recurring jobs, through their own clients and team chat. I run it daily across a VPS and a local GPU workstation.

The useful part is the knowledge around the tools. A human-readable handbook records how our systems work, while agent memory holds working context and document search supplies product facts. Keeping those separate stopped the assistant treating its own conversation history as a source of product information.

Some jobs need an agent’s judgement; others are better handled by a fixed workflow. Deciding which is which, and who can approve changes, has mattered more than adding another model.

Finding answers in 5,150 technical documents

For an industrial chemicals distributor, much of the useful product information was buried in data sheets, safety sheets and manuals. I built a search system that accepts questions in Icelandic or English and points back to the relevant document and page.

The difficult part was finding the right English source from an imperfect Icelandic query. The system searches both the original question and a translation, handles missing accents, and uses an editable glossary to correct terminology.

I chose the model that consistently supplied usable sources rather than the one that answered fastest. I also removed a ranking component when testing showed it made cross-language search worse.

The system helps people find and check a source. Retrieval tests do not establish that every generated answer is correct.

One local platform for search, creative work and internal tools

I planned a local AI workstation to support document search, image and video work, internal applications and website staging in the same environment.

The main constraint was sharing 32 GB of GPU memory. Chat and document-search models stay available for routine work; larger creative jobs queue and take over the GPU when they need it. That makes the trade-off explicit rather than letting competing workloads fail unpredictably.

Document knowledge comes from retrieval, so the source collection can change without retraining a model.

Creative freedom, without redrawing the product

A local creative studio brings image and video tools into an interface people can use without building model workflows themselves. It supports generating and editing images, restyling video and continuing a shot.

Product photography takes a different path. Existing distributor images are framed consistently, given a uniform background and upscaled only where needed. Generative models stay out of that process because they can alter packaging text or details that need to remain exact.

Rendering and source media stay local. Assistant messages use a cloud provider.

Shop assistants that prepare changes for review

Assistants can work with catalogue information, page content, SEO and analytics across two online shops, without unrestricted access to either store.

Changes to published pages become linked drafts for review. Product-feed edits have their own approval step. Each shop has a separate connection, and the service that can compare data across both is read-only.

I used existing SEO tools where they already did the job. The custom work connects those tools to the stores and makes the boundary between suggesting a change and publishing it explicit.

From asset library to marketing workbench

A full-featured digital asset management system brings creative assets, campaign planning and advertising costs into one place. Spend is tracked over time by category and medium, with a calendar for bookings, delivery deadlines and placements.

Metrics across the ads show what peaked and when. Connected online-store data puts sales and visitor trends alongside campaign activity, so the team can review what happened before planning the next campaign.

The shared view connects the material being delivered with its schedule, cost and results. Comparing those trends supports review; it does not, by itself, prove which ad caused a sale.

A catalogue people can find their way through

A B2B parts catalogue for an industrial supplier brings 3,049 products and 468 categories into a searchable WordPress site. The project is in development, with a focus on product discovery and content the business can maintain.

Site-specific search helps visitors navigate the catalogue. Custom editing blocks let the team update the front page while retaining its intended structure, rather than relying on a developer for each content change.

A separate, smaller project took a local auto-service business from a brief and logo to a live, editable website.

A traffic spike that wasn’t growth

After a tagging and consent change, GA4 showed a sharp rise in traffic. Comparing it with Search Console and commerce data showed a change in measurement, not evidence of more visitors.

That distinction mattered before anyone treated the chart as a business result. The analysis also fed a homepage brief based on search behaviour, device use and the categories visitors were looking for.