When Jev was released in September 2026, we had actually been running very small LLMs for routing and scoring, using structured outputs to pick a boolean or an enum. Jev showed the world that we can skip generating text altogether while keeping the same answer the LLM would have given.

This is exploratory work on how to build a decision head on a local LLM backbone to prove out that the mechanics work, it doesn’t require large labeled datasets, its confidence scores start matching how often it’s actually right, and training is cheap.

The hardest problem was making training reliable, since this was my first time training a model head. Gemma’s hidden states are large, so the first version of the trainable correction swung wildly from one step to the next and sometimes finished mid-swing. Reading the normalized hidden state helped with this, and a small held-out quiz now checks against overfitting and instability.

Archimedes makes investing part of everyday life with personalized lessons, playful learning, and trusted experts.

Once we launched on Product Hunt, we had massive spikes of signups where users started by playing PaperChase (think Oregon Trail for investing) where they invite their friends to play along. Through the game-play they are introduced to bite-sized lessons where they can get connected to an expert advisor's content and services.

The hardest problem was building and balancing the liquidity of a two-sided marketplace. Being a first-time founder building for both sides of the marketplace, we quickly saw that too much distribution early became a bad thing. The faster an advisor's calendar fills up, the quicker our monthly churn.

• •Acquired 10K users in first month with 16-min avg session via targeted GTM (Product Hunt, Craigslist, Facebook)
• •Launched two-sided marketplace for financial services professionals with a 20% match rate
• •Product of the Month on Product Hunt

Front-End-as-a-Service React UI delivered via API that empowers marketers to visually create stateful websites.

React was a new technology at the time. For the first version of Mason, we used plain React and the D&D kit, but it was too slow to handle big trees. We eventually had to refactor and speed up drag and drop by using plain DOM events and event propagation to reorder the elements.

Just when we thought the hardest technical problem was solved, we launched and saw that engineers hated using our tool. This led to one of the most important lessons in my career. Workflow tools need to meet the user where they do work. Engineers did not want to leave their IDEs or CLIs, but our saving grace was that marketers loved us.

• •Achieved 60% MoM user growth, 300% MoM API call growth, 200% MoM component creation
• •Sold business to Sitecore (during Covid-19)