Ask a language model whether you should take the job in another city and it will give you a considered, balanced, fluent answer. The answer has one defect, which is that no part of it came from thinking through what actually happens next. It is a well-calibrated impression of what a thoughtful response looks like.
Vernos is built on the position that the model is the wrong tool for the analysis and the right tool for the translation. The LLM turns a free-form description of a messy human situation into structured options. A simulation engine then does the reasoning, because reasoning about sequences of uncertain outcomes is a thing we have real algorithms for.
Why MCTS, and why run it twice
Monte Carlo Tree Search is the method because a life decision has the structure MCTS is built for: a branching space of paths, uncertainty at every branch, no closed-form evaluation, and a payoff only visible some distance down the tree. You cannot enumerate it and you cannot solve it analytically, so you sample it many times and let the statistics accumulate where the sampling concentrates.
The design decision I would defend hardest is that every option runs twice: a balanced base case, and a downside-weighted stress case.
This is not a UI nicety. A single expected value hides the shape of the distribution, and the shape is the entire question for decisions like these. Two options can have identical means while one has a floor you would survive and the other has a floor that ends the plan. Most people, correctly, are not optimising expectation; they are trying not to be ruined. Running an explicit pessimistic pass surfaces that directly, instead of asking someone to intuit variance from a single number.
Results are cached in Redis keyed per user, decision, option and mode, so re- examining an analysis is instant. Simulations are not cheap and people revisit these decisions repeatedly, which is what "high stakes" means behaviourally.
The part that keeps it honest
The failure mode for a product like this is producing confident output from invented inputs. A beautiful simulation of a mortgage that assumed a made-up interest rate is worse than no simulation, because it launders a guess into something that looks like analysis.
So housing decisions are enriched with real numbers. A dedicated market proxy service normalises HUD and FRED data into current mortgage rates, taxes, insurance and rent, and the simulation runs against the actual market rather than a plausible-sounding constant.
Classification has a safety net for the same reason. The LLM does the initial sort into decision types, but heuristics guarantee that the critical categories (housing, career, investment) are detected even when the model is uncertain. Those are exactly the decisions where the enrichment matters most, and they are too important to leave to a classifier having an ambiguous day. If the model is unsure whether this is a housing decision, the correct behaviour is to check rather than to shrug.
How it is put together
A Bun and Turborepo monorepo, split along the lines the workload actually imposes:
- Next.js web app: the interface where a decision is described and read.
- Express API: orchestrates the AI classification and MCTS pipeline.
- Background simulation workers on Bull queues: because a full MCTS run across every option in both modes is far too slow for a request/response cycle. Making simulation a queued job rather than a blocking call is what allows the runs to be thorough instead of budgeted to fit a timeout.
- Market proxy service: isolating the external data providers, so HUD and FRED being slow or down degrades enrichment rather than taking down analysis.
- SwiftUI iOS shell: consuming the same API, which is the return on having built an API rather than a website with endpoints.
The status says overhauling. The architecture above is sound and the thing that needs work is upstream of it: the quality of the option extraction, and how honestly the results communicate their own uncertainty. A simulation that is right about the distribution and wrong about the options is still wrong, and that is where the remaining difficulty sits.