BonaVia Compass

Personalized education should not be a luxury. Compass turns student evidence into a better-fit direction, practical projects, and a clear next step.

Comment

Inspiration

One student gets a whole think tank to plan their future. Another gets a search bar. Same potential. Very different odds.

For almost ten years, I worked in higher-education marketing, consulting, and student planning. I kept seeing the same gap: affluent families could build an entire system around one child—strategists, mentors, projects, and application support—while equally capable students often did not even know what questions to ask.

Good consulting does not really sell information. Information is everywhere. It connects the dots and turns scattered facts into a direction.

Some multi-year education-planning and application-support packages can cost six figures. The support can be valuable. The access is not fair.

I founded a registered nonprofit and built BonaVia Compass to democratize the first strategic layer of professional education planning.

What Compass does

Compass is a free bilingual education-planning experience for students and families navigating North American education.

A parent or student completes a structured Chinese or English questionnaire covering:

  • student context and current stage;
  • interests and direction clarity;
  • academic position;
  • existing experiences and outcomes;
  • goals, family priorities, and constraints.

The backend turns that evidence into a family-readable public v3/Lin report with four controlled sections:

  1. Profile — what the current evidence supports;
  2. Direction — a selected priority or clearly labelled working inference;
  3. Projects — practical ways to test that direction through real work;
  4. Timeline — a clear next step and an actionable sequence.

Compass separates what the family directly provided, what the system can reasonably infer, and what is still unknown. It does not predict admissions outcomes or choose a future for a student.

The report supports native alternate-language viewing where available and browser-based Save as PDF.

Why this is different

Most education AI answers questions or generates more recommendations. Compass is designed around a harder problem: how to make thoughtful, individualized education strategy more accessible without erasing uncertainty or student agency.

My industry experience is part of the product moat. I know which experiences help students grow, which resources create meaningful evidence for university applications, where families commonly waste time, and which questions reveal useful information about a student.

The goal is not to give families a longer activity list. It is to help them decide what is worth the student's time—and why.

Personalized education has become a luxury. BonaVia wants to make it basic infrastructure for every child.

Early family learning

Nine families used early versions of Compass.

They valued receiving, in roughly two minutes, a report that felt:

  • personal;
  • organized;
  • professional;
  • practical.

They also told us the reports were too long and needed clearer reading, more specific activities, and better resource recommendations.

Fair enough. That feedback is shaping V5: a shorter, deeper, more readable version built around verified university sources, insights from admissions professionals in my network, and structured human review.

V5 is not ready, so this submission uses the stable public v3/Lin version that I can fully stand behind.

Human review loop

Compass includes an internal review center where I can inspect the same family answers and report, add corrections, and score the result.

A delivered report is never silently rewritten. I use that feedback to improve future prompts, knowledge, evaluation rules, and tests.

What we meaningfully extended during OpenAI Build Week

Compass existed before the submission period. During Build Week, Codex and GPT-5.6 were used to recover, isolate, test, and meaningfully extend the protected public v3/Lin product.

The Build Week extension added:

  • lease-guarded report generation across worker, API, recovery, and submit-side continuity paths;
  • conditional ready and failure writes so stale workers cannot overwrite newer results;
  • stateful coverage for stale success, stale failure, and duplicate claims;
  • public-copy guards that remove internal planning markers and unresolved values;
  • conservative direction inference labelled as inference rather than a user selection;
  • a deterministic, credential-free bilingual judge preview;
  • a failing integration test that caught and removed an accidental V5 conversion from the judge path;
  • a one-command Build Week verification harness and deterministic lease-interleaving proof;
  • GitHub Actions verification.

V5 and V5.1 are not part of the submitted product or demo.

Technical implementation

The public experience uses browser JavaScript and Node.js serverless handlers. Supabase is the persistence layer in the deployed architecture. Report generation can follow an OpenAI API path or an OpenClaw-orchestrated worker path behind explicit controls and quality checks.

For judges, the reproducible path is intentionally narrower and safer. It:

  • runs on 127.0.0.1;
  • requires no credentials;
  • uses fictional data only;
  • makes zero external requests;
  • does not connect to production services.

The central reliability mechanism is a timestamped generation lease.

Imagine two workers generating the same report. If the older worker finishes last, it could erase the newer result. Compass requires a state-changing write to match the active generation lease, so late success and late failure writes are rejected instead of corrupting the final report.

The repository now includes:

  • 227/227 automated tests passing;
  • a dedicated 2/2 public v3/Lin boundary suite;
  • a complete questionnaire-to-report browser smoke test;
  • zero external requests in the judge path;
  • a credential-shaped committed-file scan with zero findings;
  • a deterministic lease-interleaving proof using the real guard functions;
  • green GitHub Actions runs on both main and the hackathon branch.

How GPT-5.6 and Codex were used

I have no coding background.

After GPT-5.6 became available, I used ChatGPT 5.6 Sol to stress-test the product plan and architecture, Codex to implement it, and ChatGPT Pro to understand bugs and design repairs.

Codex and GPT-5.6 helped me:

  • preserve and recover a messy worktree;
  • establish the latest safe v3-compatible base;
  • isolate the intended Build Week change set;
  • reason through asynchronous failure paths;
  • write failing tests before fixes;
  • detect stale-worker and duplicate-claim races;
  • catch a judge preview that still converted reports to V5;
  • build a reproducible test and verification path.

The strongest contribution from GPT-5.6 was not more copy. It was helping me reason about system states that looked correct in isolation but failed when multiple processes interacted.

Without GPT-5.6 and Codex, Compass would still be an idea in my head. These tools did not give me the mission. They gave me the ability to test it with real families.

Why this pushes AI for education forward

AI is making information easier to find. The harder question is what all that information means for one real child.

Compass does not choose a student's future. It gives the family evidence, a next step, and a chance to learn more about the student through real action.

As knowledge becomes more abundant, education must become more personal.

The long-term vision is a parallel education layer that complements schools and helps students move through a cycle of:

evidence → direction → real project → feedback → new evidence

What I learned

Building Compass changed how I think about access.

I began by asking: Who gets good guidance?

Then I realized there was another question: Who gets to build?

When more people can build, more problems can be solved by the people who understand them best.

That is how progress compounds.

Limitations

  • The submitted judge path is local and synthetic; it is not a production-scale claim.
  • It does not predict admissions outcomes or guarantee educational results.
  • The offline preview does not exercise production persistence, delivery, or remote model calls.
  • Only fictional data is used in the judge flow.

What's next

V5 is being rebuilt around family feedback, verified university information, admissions expertise, clearer reading hierarchy, and more specific activity and resource recommendations.

The goal remains simple: help more students find a path that truly fits them.

Built With

  • browser-print-api
  • github-actions
  • gpt-5.6-sol
  • html/css
  • javascript
  • node.js-22
  • openai-api
  • openai-codex
  • openclaw
  • supabase
  • zod
Share this project:

Updates