ExamVault

Inspiration

I am a physician preparing for my cardiology board examination. The amount of material is overwhelming: extensive textbooks, constantly evolving guidelines, and years of exam questions that often have no official answers. For a highly specialized examination like this, there are far fewer comprehensive question banks and broadly accepted explanations than for broad standardized medical examinations such as the USMLE.

Traditionally, I had only two options: investigate every question across multiple sources by myself, or collaborate with other physicians and share our reasoning. Both approaches are valuable, but they are difficult to maintain. When a guideline changes or an important study is published, an answer that was once accepted may become outdated—and it is often unclear which source originally supported it.

Reading the original textbooks, guidelines, and research papers remains essential. Yet within hundreds of pages of detail, it is easy to lose sight of the specific distinctions that actually determine an exam answer.

I began to imagine each exam question as a star placed back onto the original text. One question is only a point, but when many tested points are connected in the source's own conceptual order, they begin to form a constellation. That constellation could reveal the most important patterns without removing them from their original context.

This idea became ExamVault: a system that connects exam questions to exact-page evidence, organizes tested concepts back into the structure of the source, and helps learners find what matters with less searching and less cognitive friction.

What it does

ExamVault turns assigned textbooks, guidelines, and past exam questions into an auditable learning loop.

Each question becomes an Ultra Card containing the smallest decision key, option-by-option reasoning, and links to the exact supporting PDF pages. Tested concepts from multiple questions are then reorganized into an Ultra Note following the source's original conceptual order. Finally, the Reader overlay highlights where each tested point appears within the complete source text.

This point-to-line-to-surface structure helps learners identify high-yield distinctions quickly without separating them from their original context or evidence.

How I built it

ExamVault is implemented as an Obsidian vault with Python validation and indexing tools, two project-owned JavaScript plugins, Markdown-based learning artifacts, and a local SQLite FTS5 page index.

Codex with GPT-5.6 performs source-bounded interpretation, decision-key extraction, option adjudication, exception handling, and source-ordered distillation. Deterministic tools handle PDF intake, page identity, content hashes, local indexing, schema validation, privacy checks, and release gates.

The bundled demo requires no API key, hosted account, or network connection. It includes only reviewed, openly licensed or project-authored educational materials.

How I used Codex and GPT-5.6

ExamVault existed before Build Week, but Codex with GPT-5.6 transformed it into a focused, testable competition release.

Codex inspected the repository, helped consolidate the workflow into one active production route, implemented and tested the judge-facing experience, created synthetic scientific fixtures, strengthened the Obsidian plugins, and built deterministic privacy and release validators.

GPT-5.6 was especially valuable when reasoning had to remain inside an assigned evidence boundary: identifying the decisive distinction, comparing every answer option, preserving exceptions, and reorganizing tested concepts without losing their relationship to the original source.

The project owner made the final product, evidence-authority, privacy, licensing, and release decisions.

Challenges I ran into

The hardest challenge was not generating explanations—it was proving that every explanation came from the correct source and exact page.

I also had to preserve stable links while transforming source material into readable notes, support questions with different numbers of answer choices, distinguish model reasoning from deterministic validation, and create a judge-ready demo without exposing private study materials, copyrighted textbooks, patient data, or secrets.

Accomplishments that I'm proud of

I created a complete evidence loop in which a learner can answer a question, reveal its reasoning, inspect every option, open the exact supporting PDF page, review the concept in source order, and see the tested passage highlighted inside the Reader.

The release includes four validated demonstration cards, exact-page evidence links, source-ordered Ultra Notes, auditable Reader overlays, local page-level retrieval, and automated release checks. The entire judge path runs locally without patient data or network access.

What I learned

AI becomes substantially more useful for specialized education when it operates inside an explicit evidence boundary.

A fluent answer is not enough. Learners need to know which source supports it, where the supporting passage appears, how competing options were rejected, and whether uncertainty remains unresolved. I also learned that model reasoning and deterministic validation are complementary rather than interchangeable.

What's next for ExamVault

Next, I plan to improve OCR and CJK retrieval, support more scientific specialties, detect changes between guideline editions, and make collaborative question review easier.

The long-term goal is to help learners build living, updateable maps of what has been tested—while keeping every conclusion connected to its original evidence.

Built With

• codex
• fts5
• gpt-5.6
• javascript
• json
• markdown
• obsidian
• openai
• poppler
• python
• sqlite