MemoryBun

Inspiration

The Oxbridge interview is the final, high-stakes hurdle for prospective students—where they only have one or two chance to showcase years of intellectual potential. However, success depends not solely on ability but also on access to elite interview preparation. While students at top-tier schools receive hours of expert coaching, others face cold-outreach rejections or high-quality mock sessions costing £20+ per hour.

MemoryBun was created to dismantle these barriers. We provide an AI-powered mock interview tutor that delivers the personalized feedback previously gated behind a paywall. While AI may not fully replace a human interviewer yet, it offers an immediate, high-value alternative that is 10x cheaper—or even free. Grounded in the principles of open education, our mission is to ensure that the brightest minds, not just the best-funded ones, have a seat at the table.

What it does

MemoryBun is an AI-powered platform that simulates the high-pressure Oxbridge technical interview.

Realistic Simulation: Practice multi-part technical problems with timed prep and recording phases.
Interactive Whiteboard: Work through complex solutions and diagrams on an integrated canvas.
Live Transcription: Capture thought processes in real-time with speech-to-text as you speak your answers.
AI-Driven Grading: Receive instant scores across five key dimensions: Problem Framing, Solution Execution, Technical Correctness, Communication, and Time Management.
Performance Analytics: View actionable insights and radar charts to visualize strengths and track improvement over time.

How we built it

Architecture Overview

MemoryBun utilizes a high-performance, event-driven microservices architecture designed to minimize latency and handle complex AI workloads.

Frontend: Built with TypeScript, React, and Vite. It features a custom HTML5 Canvas for whiteboarding, KaTeX for mathematical rendering, and leverages the MediaRecorder API for high-fidelity audio capture.
Backend: A lean FastAPI ecosystem managing three core services:
- Question Service: Orchestrates CRUD operations for questions, rubrics, and metadata.
- Transcription Service: Utilizes faster-whisper workers to process audio chunks via Redis queues.
- Grading Service: Employs an orchestrator pattern—TaskDecode → ContextFetch → PromptBuild → LLMGrade (Gemini 3 Flash) → Validate—to provide structured, modular feedback.

Challenges We Ran Into

1. The "Post-Interview Wait" (Latency)

Initially, students faced an unacceptable 10+ second delay after finishing their response. We solved this by implementing a streaming chunking strategy, splitting audio into 30-second segments processed mid-recording. By switching to faster-whisper (CTranslate2 optimized) with int8 quantization, we slashed memory overhead and achieved near-instant results.

2. Reliable Structured JSON Output

Forcing an LLM to follow a strict grading rubric often results in a trade-off: speed vs. structural integrity. We optimized this by:

Prompt Engineering: Disabling Chain-of-Thought (CoT) for simple tasks to reduce token latency.
Model Tiering: Using Gemini 3 Flash for multimodal inputs (whiteboard + audio) and a lighter variant for text-only summaries.
Validation: Implementing a robust "Validate → Persist" pipeline to catch and fix malformed JSON before it reaches the UI.

3. Designing Feedback That Matches Real Oxbridge Interviews

Oxbridge interviews are not fully transparent—there is no public marking scheme, and success is not about arriving at the “right” answer. To address this, we conducted extensive research and spoke with current students and alumni to understand how candidates are actually assessed. A clear pattern emerged: interviewers evaluate thinking, not outcomes.

Our AI rubrics are therefore designed to prioritise reasoning and structure over final correctness, ensuring feedback reflects what Oxbridge tutors genuinely look for rather than common misconceptions.

Accomplishments We’re Proud Of

Ultra-Lean Infrastructure: The entire stack runs on a single T3.micro. We proved that a high-impact AI tool doesn't require heavy, expensive infrastructure.
Elegant Simplicity: Built without a persistent database—using JSON for content and Redis for transient states—allowing for lightning-fast iteration.
Sub-3s Response Times: Final evaluations are delivered within ~3 seconds of the recording ending.
True Multimodal Grading: Our AI doesn't just "hear" the student; it "sees" the whiteboard, grading the synergy between verbal explanation and visual problem-solving.
Premium UX: A responsive, "Stitch-inspired" interface that feels like a polished professional tool, not "AI slop."

What's Next for MemoryBun

Subject Expansion: Scaling the question bank to cover a wider range of STEM and Humanities subjects with varying difficulty tiers.
Alumni-Verified Content: Partnering with Oxbridge alumni to source and verify authentic, high-fidelity interview question, build a community
Dynamic Follow-ups: Implementing ad-hoc AI follow-up questions to better simulate the "prodding" nature of a real academic interview.
End-to-End Productisation: Designing and launching a user panel with progress tracking, session history, and personalised insights to support sustained practice beyond single mock interviews.