<aside> 🎓
PaperTrial — Complete Technical Overview
The definitive technical document detailing every implemented feature, pipeline, and architectural flow in the PaperTrial ecosystem.
</aside>
Live : https://paper-trial.vercel.app/
PaperTrial is a fully functional, AI-powered academic thesis defense simulator.
Defending a thesis is a high-anxiety bottleneck for university students. PaperTrial acts as a hyper-realistic, adversarial AI professor that ingests the user's PDF thesis, finds methodological weaknesses, and grills them verbally in real time.
Unlike generic chatbots, PaperTrial forces the user into a high-pressure, voice-only live environment while tracking academic integrity and grading spoken answers against the exact text of the uploaded thesis.
Every single feature actively implemented and running in the current PaperTrial ecosystem.
| # | Feature | Description |
|---|---|---|
| F1 | Next.js Clerk Authentication | Secure login and signup portal via Clerk UI. |
| F2 | Backend JWT Validation | FastAPI decodes JWTs to enforce strict row-level security for user data. |
| F3 | Drag & Drop Upload UI | React dropzone for securely uploading PDF thesis files. |
| # | Feature | Description |
|---|---|---|
| F4 | PyMuPDF Text Extraction | Uses fitz.open() to extract raw text, identify paragraphs, and preserve page metadata. |
| F5 | Live SSE Progress Tracking | Backend streams JSON updates via Server-Sent Events to a live UI progress bar. |
| F6 | Semantic Text Chunking | LangChain RecursiveCharacterTextSplitter divides text into ~500-character chunks with overlap. |
| F7 | Google Vertex AI Embedding | text-embedding-004 converts chunks into 768-dimensional dense vectors. |
| F8 | PostgreSQL pgvector Storage | SQLAlchemy stores vectors in an HNSW-indexed pgvector database for fast retrieval. |
| # | Feature | Description |
|---|---|---|
| F9 | Structural Flaw Detection | Gemini 2.5 Flash analyzes chunks to detect sampling bias, logic gaps, and methodology issues. |
| F10 | Attack Question Synthesis | Gemini 2.0 Flash condenses weaknesses into exactly 5 targeted attack questions. |
| F11 | Thesis Health Scoring | Aggregate pre-defense score across 5 dimensions. |
| F12 | Pre-Defense Radar Chart | Recharts radar visualization of the health dimensions. |
| # | Feature | Description |
|---|---|---|
| F13 | Session Initialization | FastAPI manages session state and loads the first attack question into the defense queue. |
| F14 | Real-time Mic Streaming | Frontend uses MediaRecorder to stream audio bytes over secure WebSockets. |
| F15 | Deepgram STT | Low-latency speech-to-text transcription of spoken answers. |
| F16 | Voice Metric Tracking | Real-time Words Per Minute measurement to estimate confidence and hesitation. |
| F17 | Academic Integrity Tracker | visibilitychange listener deducts points if the user switches tabs to reference notes. |
| F18 | RAG Context Retrieval | Embeds the spoken transcript and retrieves top-3 relevant thesis segments via cosine similarity. |
| F19 | Real-time Gemini Grading | Grades transcript vs retrieved context, scoring accuracy, depth, and coherence. |
| F20 | ElevenLabs Voice Synthesis | Generates an authoritative AI professor voice through TTS. |