Jevfish: Our CTO Built a Chess Engine That Never Calculates

Our CTO, Pate Bryant, built a chess engine called Jevfish that beats humans without evaluating a single position tree. It is a clean demonstration of a software architecture pattern that we apply for clients every day.

The Engine That Never Calculates

Traditional chess engines like Stockfish search millions of positions per second, building enormous trees of future board states to find optimal moves. Jevfish does none of that.

The whole engine is a few hundred lines of code. Here is how it works:

1. Generate all legally valid moves for the current board position
2. Filter that list down to tactically sound options
3. Pass the filtered list to a decision model and ask: which of these?

The model answers in under a second. No deep search. No position evaluation functions. Just a fast classification call over a bounded, rule-constrained option set. Jevfish went 3-0 against human players, including Pate himself.

Classification, Not Generation

This is the architectural insight at the core of Jevfish. Most AI integrations use language models to generate free-form text. Pate used a decision model to classify among a constrained set of valid options.

The difference matters in real software:

• Typed output - the model returns a specific move, not a paragraph describing one
• Speed - constrained classification runs in milliseconds
• Cost - far cheaper per call than open-ended text generation
• Reliability - code owns rule enforcement; the model owns judgment

The code and the model each do what they are best at. Code enforces the rules and generates the legal option set. The model handles the judgment call. Neither does the other's job. Together, the system plays real chess at a level that beats the person who built it.

Why This Pattern Matters Beyond Chess

Chess is a clean domain to demonstrate a pattern that applies across software: when you have a well-defined, bounded decision space, a decision model outperforms exhaustive search on cost, speed, and maintainability.

The same architecture fits anywhere your software needs to make fast choices from a bounded option set: workflow routing, prioritization queues, recommendation systems, branching logic, triage, and scheduling. You do not need a large generative model running expensive inference. You need a classifier that answers "which of these?" in a tight loop.

Pate built Jevfish to make this concrete and demonstrate it against a real opponent with objective win/loss outcomes. The result: 3-0, playing correct fundamentals, in a few hundred lines of code.

Pate's Breakdown

Pate shared the full technical breakdown on his LinkedIn, including how Jevfish handles uncertain positions (a two-move lookahead that still uses the model's judgment rather than a search tree) and how it converts endgames.

If you are building software that needs to make fast, reliable decisions over a constrained set of options, this is the architecture to understand. Talk to our team about how it applies to your stack.