AI Enterprise SolutionsGPT & Claude

BrightLine Academy — AI exam-prep platform for Azerbaijan's national exams

Users: school students preparing for Azerbaijan's state exams run by DİM (graduation exams, university admission groups I–V and master's admission), their teachers, and their parents. Problem: students practise from a few printed books whose questions run out quickly, have no step-by-step explanations, and give no feedback on weak topics. Private tutoring is expensive. General chatbots answer in Turkish or English, do not follow the DİM syllabus, and often give a wrong answer key. Outcome: unlimited DİM-style practice questions (A–E) in correct Azerbaijani, each with a step-by-step explanation. Every AI-generated question is checked automatically before the student sees it. The platform also offers an AI tutor, a timed exam simulator with DİM scoring, weak-topic analysis, a real-time quiz, and a teacher panel.

Open the demo Video GitHub

Preliminary score

69.5/ 100Promisingaverage of 2 judges

Value for the user18.0 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI23.0 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing15.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility7.0 / 15
Data requirements, running costs and a clear next step.
Originality6.0 / 10
A meaningful difference from common solutions.

What the judges wrote

GPT-6 AstraPreliminary71/100
Strengths
Your DİM-specific syllabus, Azerbaijani explanations and exam-group selection address concrete limitations of printed practice books and generic chatbots. Your screenshots and implementation details support a substantive prototype where AI contributes directly to practice generation and tutoring. Your separation of stubbed end-to-end tests from live Gemini checks, together with specific fixes for answer-key mismatches, formula rendering and real-time privacy leaks, is particularly strong.
To improve
Your next evaluation should use teacher-reviewed questions to measure factual correctness and syllabus alignment against a generic chatbot; passing structural and consistency checks does not establish educational accuracy. Show one complete generation → attempt → explanation → saved-progress run, including rejection of an invalid question. Your feasibility plan needs API cost and latency per accepted question, including regeneration, plus a focused pilot measuring learning gains or teacher preparation time saved. Enable HTTPS before collecting student data and reconcile the deck's conflicting statements about whether teachers can see student phone numbers.
Claude Opus 5.5Preliminary68/100
Strengths
Your AI question generator does real work: it produces DİM-format A–E questions in Azerbaijani and then validates each one automatically. It checks for unique options, an answer key that matches the explanation, balanced LaTeX and no Turkish words, and regenerates questions that fail. Your failure list is unusually honest and specific: wrong keys, geometry figures leaking solution values, logic questions truncated at the token limit, and opponents' answers leaking over SSE, each with a fix and a regression test. The live demo, the 34 end-to-end tests with a stub AI, and the 80-question live Gemini run with input→output logs give judges real evidence rather than claims.
To improve
Your quality checks are structural, so a question can pass all of them and still be mathematically wrong or off-syllabus. Have a teacher review a sample, or cross-check maths answers with a solver, and report the error rate. Your comparison with today's approach stops at 'printed books have fewer questions'. Measure something like time-to-explanation or score gains for a few real students. The materials never mention running costs (Gemini cost per student per month) or a concrete next step such as a school pilot, pricing or a partner, which the Enterprise track explicitly asks for as an adoption metric. Fewer features and more depth on the core generate→validate→learn loop would make the value case stronger.

The team

Soltan RzayevLead
product
Nihat Damirov
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Google Gemini (gemini-3.8-flash), called through an OpenAI-compatible chat API in JSON mode. It is used for question generation, the AI tutor, image-based problem solving, mini-lessons, weak-topic analysis, the daily plan and the CV builder. Voice input and output use the browser's built-in Web Speech API, not an AI model. We did not train or fine-tune any model.
Data: No external training dataset. Questions are generated by the model from a DİM syllabus topic list written into our prompts (backend/.../ai/AiPrompts.java). University cut-off scores were imported from public admission data. Admins can upload real past exam documents, and the AI extracts questions from them. User data is stored in PostgreSQL.
Components: Frontend: Next.js 16, React, TypeScript, Tailwind CSS, shadcn/ui, KaTeX (formula rendering), PWA service worker. Backend: Spring Boot 3.3, Java 21, Spring Data JPA, PostgreSQL (production), H2 (tests only), Jackson, Server-Sent Events for real-time features. Testing: JUnit 5, Spring Boot Test. All are open-source libraries; no paid templates.
Built during the event: Confirmed

Quality testing

34 automated tests, all passing. The end-to-end tests start the real server on an HTTP port with an in-memory database and a stub AI. They cover teacher approval, private messages, unread counters, the question generator, the AI tutor, the real-time quiz and real-time privacy. Live test with the real Gemini API: 80 questions across 13 subjects/groups plus logarithms, trigonometry and geometry. Each question is checked automatically: 5 unique options, the answer key matches the explanation, balanced LaTeX, no Turkish words. Result: 80/80 questions passed. KaTeX rendered all 717 formulas with 0 errors. What broke and how we fixed it: 1. The AI calculated the right answer in the explanation but marked a different option as correct. Now such questions are rejected and regenerated. 2. Unclosed formulas ($...) garbled the text. Now rejected. 3. Geometry figures revealed solution values. Values not given in the question are now hidden ("?"). 4. Logic questions came back empty because they were cut off at the token limit. Fixed with JSON mode and a higher limit. 5. In the real-time quiz, a player could see the opponent's answers. Updates are now neutral; a test reads the live SSE stream to confirm it (group chat had the same issue and is fixed too). 6. Vector formulas were split and shown as raw text. Fixed in the formula renderer. Comparison with the current approach: printed books have a fixed, small question set and no explanations. We generate unlimited questions, and each one is validated before a student sees it. Full input → output log: docs/test-neticeleri/ in the repo.

Worth 20 of the 100 points.

Running it

Demo: open http://13.60.44.113/ → "Qeydiyyat" → choose "Şagird" (Student). No approval is needed. The browser shows "Not secure" because the site uses HTTP; voice input needs HTTPS, and everything else works. Try: Sual Generatoru (pick a subject, or "Təsadüfi" for a random subject) → AI Müəllim → İmtahan Simulyatoru → Bilik Arenası. Tests (no API keys or database needed; they use an in-memory H2 database and a stub AI and never touch production data): cd backend && ./gradlew test Live AI quality test (calls the real Gemini API): AI_LIVE_TEST=1 GEMINI_API_KEY=... ./gradlew cleanTest test --tests '*LiveAiQualityTest' Local run: copy backend/src/main/resources/application.yml.example to application.yml, then set PostgreSQL and the Gemini key (or use the "h2" profile). cd backend && ./gradlew bootRun npm install && npm run dev Production runs on PostgreSQL. Test results (input → output): docs/test-neticeleri/

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.