PaperYak
AI-powered essay grading platform built for UK universities. Next.js frontend, Claude API backend, processing 2,500-word essays in under 60 seconds.
The Challenge
UK university students receive essay feedback weeks after submission. Hiring private tutors costs £40-80/hour. Generic AI tools like ChatGPT don't understand UK academic standards (First, 2:1, 2:2, Third classifications) or proper citation formats (Harvard, OSCOLA).
We built PaperYak to provide instant, detailed feedback calibrated to UK marking criteria. The system needed to handle everything from structure and argumentation to citation checking and grammar - all while matching the quality a human tutor would provide.
Complex Technical Challenges
Building an AI grading system required solving problems most platforms ignore
UK vs US Academic Systems
The Problem: UK universities use entirely different grading scales (70+ = First, 60-69 = 2:1) and expect different essay structures compared to US institutions. Generic AI models trained on American data fail completely when evaluating British essays.
Our Solution: We built a dual-context prompt engineering system that detects writing patterns (spelling: "colour" vs "color", citation style, essay structure) to automatically switch between UK and US evaluation frameworks. This required creating separate rubric mappings for each system and training the AI on authentic UK university marking criteria sourced from Russell Group universities.
Multi-Style Citation Detection
The Problem: Students use Harvard, APA, MLA, OSCOLA, and Chicago styles - often mixing them incorrectly. Detecting missing citations requires understanding context: is a claim common knowledge or does it need a source?
Our Solution: Hybrid detection system combining regex pattern matching (for format errors) with AI semantic analysis (for missing citations). We built a custom parser that identifies in-text citations, validates them against bibliography entries, and flags unsupported claims. Trained on 200+ annotated sample essays to reduce false positives.
Sub-60-Second Processing at Scale
The Problem: Students expect instant feedback, but Claude API calls take 10-20 seconds. Add citation parsing, grade calculation, and report generation, and you're looking at 2+ minute wait times - unacceptable for modern UX.
Our Solution: Optimized prompt engineering to reduce token count by 40% without losing accuracy. Implemented parallel processing (citation check runs while AI analyzes content). Added streaming responses so users see feedback appearing in real-time, not after waiting for the full analysis. Result: 2,500-word essays graded in under 60 seconds.
Academic Standard Calibration
The Problem: A "First" at Oxford is different from a "First" at a teaching-focused university. Generic AI either over-grades (to please users) or under-grades (to seem rigorous). Neither helps students improve.
Our Solution: Worked with academic staff to define objective criteria for each grade band. Built evaluation matrices covering 12 dimensions (argument quality, evidence use, structure, critical analysis, etc.). System provides grade ranges rather than absolutes ("likely 2:1, could reach First with stronger conclusion") to match real academic feedback.
Technical Stack
Frontend
- Next.js 14
- React 18
- TypeScript
- Tailwind CSS
Backend
- Claude API
- Clerk Auth
- Stripe
- PostgreSQL
Infrastructure
- Vercel
- Edge Functions
- Neon DB
- Resend
Live & Growing
PaperYak is now live, helping UK students get instant feedback on essays
Need a Platform Built?
We build clean, functional web applications - from AI platforms to e-commerce sites
Let's Talk About Your Project