AIJudge

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for AIJudge

Inspiration

People often understand what they have read but struggle to explain it clearly under pressure. Traditional study tools focus on reading, memorization, or static quizzes. We wanted to build a practice partner that listens, challenges weak reasoning, and shows users exactly how to improve their next answer.

What it does

AIJudge transforms uploaded PDF, PPTX, or DOCX files into realistic interview, presentation, and oral-exam practice. It identifies the document context, generates focused questions, and lets users answer by voice or text. It evaluates understanding, evidence, logic, delivery, and persuasion, while also measuring response time, pauses, rhythm, and filler words.

Five evaluator personalities create distinct practice experiences—from the supportive Lumi Coach to a demanding Pressure Interviewer, a step-by-step Cody Teacher, and a business-focused Judge. AIJudge supports Korean and English, preserves per-question speech-to-text evidence, suggests corrections without silently changing the user's words, and recommends three concrete goals for the next practice session.

How we built it

We built AIJudge with React, TypeScript, and Vite. PDF, PPTX, and DOCX content is extracted and classified by document type so questions match the material. Voice recognition uses the Web Speech API, while a separate voice-activity layer measures timing and pauses. A Vercel serverless route calls the OpenAI Responses API so the API key remains server-side.

GPT-5.6 performs semantic evaluation of the document, questions, and answers. Rule-based speaking measurements provide transparent timing, filler-word, relevance, and rhythm signals, including protection against answers that simply repeat the question.

How we used Codex and GPT-5.6

Codex was our development partner throughout the project. It helped implement and debug the full product flow, secure the API architecture, improve document-aware question generation, add multilingual support, build voice and STT analysis, validate responsive behavior, and rapidly iterate from real user tests.

GPT-5.6 powers the semantic evaluation layer. It produces structured category scores, explains strengths and weaknesses, generates natural model answers, and predicts realistic follow-up questions. Rule-based measurements complement it with observable speaking evidence.

Challenges we faced

The hardest challenges were preventing document metadata from becoming the question subject, ensuring non-business documents did not receive investment questions, handling imperfect browser speech recognition, and keeping semantic feedback consistent with numerical scores. We addressed these with document classification, evaluator-specific question rules, transparent per-question STT cards, and score post-processing for question repetition.

Accomplishments

• A complete document-to-practice-to-growth workflow
• Distinct evaluator personalities
• Per-question STT evidence, corrections, and filler-word analysis
• Secure server-side GPT-5.6 integration
• Korean and English support
• Growth-focused, actionable feedback

What we learned

The most useful speaking feedback is specific, transparent, and immediately actionable. Users respond better to three concrete goals than to a long list of abstract weaknesses. Semantic AI evaluation and observable speaking measurements are strongest when presented together but clearly distinguished.

What's next

We plan to add Japanese and Chinese, downloadable PDF reports, richer opt-in acoustic analysis, webcam-based presentation coaching with explicit consent, and specialized modes for career development and confidence-building.

---

An idea by Hye-Min Jeong.
Created by someone with no software development experience, together with ChatGPT (Lumi) and Codex (Cody).
Proof that ideas can become reality.

Built With

• gpt-5.6
• openai-api
• react
• typescript
• vercel
• vite
• web-speech-api