Compass

Inspiration

Building a project today rarely means making decisions in isolation. A single idea can involve hundreds of choices about architecture, APIs, models, data, scope, security, deployment, and product direction.

The problem is that the reasoning behind those choices is usually scattered across AI chats, documentation, repositories, notes, and conversations. As a project evolves, builders often have to reconstruct why something was decided before they can decide what to do next.

This led us to a simple thesis:

Decision is infrastructure.

I wanted to build something that would preserve the reasoning surrounding a project—not just generate another answer—and help builders navigate uncertainty without repeatedly losing their context.

That became Compass.

What it does

Compass is an AI reasoning layer for builders that helps them understand and navigate complex project decisions.

The current prototype focuses on:

AI Project Navigation — helps builders reason through complex decision branches and determine practical next steps.
Feasibility Assessment — evaluates proposed approaches against technical and project constraints.
Structured Analysis — breaks ideas and decisions into their underlying technical primitives.
Assumption Detection — identifies implicit assumptions and dependencies that could become problems later.
Risk Identification — surfaces structural failure modes before they become expensive to fix.
MVP & Journey Mapping — transforms decisions into an actionable roadmap and helps visualize how the project can evolve.

Instead of treating every interaction as an isolated AI prompt, Compass is designed around the project's evolving decision context.

How we built it

I designed Compass around a model-agnostic reasoning architecture so that its core logic is not tied to a single AI provider.

The conceptual architecture is:

Builder Intent / Decision → Evidence & Reasoning → Assumption Dependency Map

with the reasoning layer connected to future capabilities such as repository intelligence and an anticipation engine.

For the prototype, I deliberately focused on the parts that could provide immediate value to a builder rather than attempting to implement the entire long-term architecture.

I structured the system around project context, decisions, assumptions, constraints, risks, and execution paths. The AI then uses that structured context to provide reasoning and navigation rather than simply producing generic text.

Challenges we ran into

The biggest challenge was scope.

Compass was originally envisioned as a much larger decision infrastructure system, including confidence measurement, evidence tracking, predictive context, repository intelligence, collaboration, and multi-perspective review.

With limited development time, implementing all of these properly would have resulted in a much less reliable prototype.

We therefore had to make difficult decisions about what belonged in the first version. The Confidence Meter, for example, was moved from the prototype into the future roadmap so that I could concentrate on the core reasoning experience.

Another challenge was making the AI reasoning useful without turning Compass into another generic chatbot. The system needed to focus on relationships between decisions, assumptions, risks, and constraints rather than simply responding to prompts.

Accomplishments that we're proud of

I are proud that Compass evolved beyond the idea of "AI that gives advice" into a clearer concept: AI as persistent decision infrastructure.

I created a prototype that focuses on a problem that is easy to overlook but increasingly important as AI accelerates software development: maintaining continuity of reasoning while a project changes.

We are also proud of the architectural direction. By keeping the reasoning layer model-agnostic, Compass can potentially evolve alongside different AI models and tools instead of being permanently coupled to one provider.

Most importantly, I learned to prioritize a coherent, usable core over implementing every feature in the original vision.

What we learned

I learned that the hardest part of building an AI product is not always getting a model to generate an answer.

The harder problem can be determining what context the model should reason over, which assumptions should be challenged, and how previous decisions should influence future ones.

I also learned that good AI products need explicit boundaries. Not every ambitious feature belongs in the first prototype. Defining the smallest useful version of a system can be as important as designing its eventual architecture.

What's next for Compass

Compass is intended to grow from an AI reasoning assistant into a more complete decision infrastructure for builders.

Our roadmap includes:

Confidence & Evidence Engine — including the deferred Confidence Meter, Evidence Ledger, and richer assumption tracking.
Predictive Context Layer — an Anticipation Engine that identifies likely downstream bottlenecks and consequences.
Evolution Timeline — preserving how and why a project changed over time.
Repository Intelligence — connecting decisions and assumptions directly with the project's codebase.
Multi-Perspective Review — evaluating decisions from Engineering, Security, Business, and other perspectives.
Team Collaboration & Audit Trail — making project reasoning understandable and traceable across teams.
Model-Agnostic Architecture — allowing Compass to work with evolving AI models rather than depending on a single model provider.

The long-term goal is simple:

Help builders spend less time questioning where they are going and more time building what they believe should exist.