AI product builder

Thor
Matthiasson

AI products grounded in real business problems.

My background is in performance marketing and analytics. My work explores how AI products can support real workflows, and how to tell whether their outputs are useful.

Get in touchGitHubLinkedIn
Thor Matthiasson
Thor Matthiasson AI With Thor

Two systems to inspect.

Publishing agent-built applications and making model disagreement reviewable.

Live product · Beta

Pagenta

From an AI-generated artifact to a usable application.

Product stack: TypeScript, Next.js, Supabase/Postgres, REST, CLI and MCP.

Visit Pagenta

The problem

An agent can generate an interface, but someone still has to publish it, control access, save state and keep updates working. That handoff is part of the product.

What the product does

Pagenta gives agents and people a publishing workflow: create a page, deploy it to a stable URL, then revise the same application. The platform includes persistent state and access controls, with dashboard, CLI and MCP entry points.

What makes the workflow useful

The result needs to work for someone outside the AI conversation. A stable link, access controls and a way to revise the page make the generated artifact easier to share and use.

A workflow to explore

Publish through MCP, read the page back, update it at the same URL, and check the result in a browser. The product question is whether the artifact survives the handoff to a real user.

Personal MCP · Open source

Plurality

Make different models' reasoning available for review.

Product stack: TypeScript, Nuxt, model-provider integrations, SQLite and MCP.

Visit PluralityRead the MCP source

The problem

A polished answer hides uncertainty. A useful review needs to show the competing arguments, the assumptions they share and the evidence that could change the decision.

What the product does

Plurality orchestrates models through eight methods, including debate, red teaming, peer review and forecasting. Its personal MCP server lets an agent plan a run, start it, inspect its status, retrieve the result and export the record.

A tradeoff to keep visible

More model calls add cost and latency. Models can share the same blind spots, and several personas on one model are not independent evidence. Agreement needs to be assessed against sources and a single-model baseline.

A workflow to explore

Run a structured review, inspect individual contributions, recover a failed run and export its saved history. The public MCP repository provides code and setup instructions for the personal version.

When the evaluation needs evaluating.

A research investigation into apparent collusion between LLM pricing agents.

A rival cuts its price after being undercut. Does that demonstrate a collusive punishment strategy?

My pricing-agent study examines 240 simulated markets, with five models pricing for 300 periods per market. The experiment replays stored market states and compares forced price moves with matched controls and Q-learning agents.

The failure was in the inference: a positive forced-cut test could be read as evidence of collusion even when a simpler explanation fit the behavior.

What this means for AI products

Test the metric against a simpler baseline. Change the intervention. Look for cases where the test passes for the wrong reason. Revise the claim when the controls contradict it.

The initial signal

In high-gain markets, 84% of rivals cut by more than 1% after a forced cut, compared with 8% between control branches.

The control that matters

A rival following only the shift in its one-shot best response would clear that same threshold in 98% of those markets. The threshold alone cannot identify collusion.

The revised conclusion

Responses were roughly proportional, followed price rises too, and stayed lower after a cut. In this simulation, the pattern was better explained by price matching than by a punish-and-return strategy.

Price Matching, Not Stick-and-Carrot

How LLM Pricing Agents Respond to Forced Price Cuts

Thor Matthiasson · Research manuscript · November 2026

Read the manuscript (PDF)

Business context.
Practical AI.

My background is in performance marketing, analytics and business operations. My AI projects give me a way to explore practical workflows and question the results. The starting point is the user's problem and the outcome they need.

I also help organize AI Tinkerers in Reykjavík and share applied AI work through AI With Thor.

I'm interested in work that connects AI capabilities with useful business workflows and clear evaluation. The projects above give us concrete products, questions and tradeoffs to discuss.

Let's talk through the work.

AI products, practical workflows and adoption.

Your details are used to reply to this inquiry.

Thor Matthiasson
GitHubLinkedInFull portfolio