Cuemath AI Tutor Screener

This project is an AI-powered voice interviewer designed for Cuemath's tutor screening process. It uses a custom-tuned LLM and browser-native speech APIs to conduct an automated, 5-minute initial voice screening, producing a quantitative assessment report against a behavioral rubric.

Design Decisions

Building this prototype required balancing user experience, latency, and system complexity. Here is the rationale behind the major architectural decisions:

1. Web Speech API over Whisper

Relying on the browser-native Web Speech API (via SpeechRecognition and SpeechSynthesis) instead of an external STT/TTS service like OpenAI's Whisper or ElevenLabs was a conscious choice for the MVP.

• Latency: Audio processing happens locally in real-time, removing network roundtrips for audio chunks. The time-to-first-byte (TTFB) for the AI's spoken response is simply the duration of the LLM text completion.
• Cost & Complexity: It removes the need for WebSocket streaming, audio chunking logic, and heavy backend infrastructure.
• Privacy (DPDP Act): Audio never leaves the candidate's browser. Only the generated text transcript is sent to the backend for analysis, strictly adhering to data protection laws by minimizing PII exposure.

2. Stateless Backend (No DB)

The FastApi backend operates entirely without a database.

• Simplicity: By persisting the session state (transcript and report JSON) entirely on the client side (localStorage) during the MVP phase, we avoid the overhead of PostgreSQL/Redis setups, user authentication, and state synchronization.
• Shareability: Reports are keyed to a generated UUID and hydrated locally, keeping the focus entirely on validating the core interaction loop rather than building generic CRUD boilerplate.

3. Exactly 7 Turns

The interview flow is strictly capped at 7 conversation turns (Priya speaks → Candidate replies).

• Structure over vibes: A fixed progression guarantees that Priya covers the required diagnostic surface area (Opener → Simplification → Patience with mistakes → Values alignment → Closing). It ensures no candidate is unfairly evaluated due to an unstructured AI wandering off-topic.
• Duration Control: 7 turns naturally equate to a ~5-minute interview, which hits the sweet spot for a low-friction top-of-funnel screening process.

4. BARS-Anchored Rubric (Behaviorally Anchored Rating Scale)

Instead of asking the LLM to simply "score warmth from 1 to 5," the ASSESSMENT_SYSTEM_PROMPT embeds highly specific, behavioral examples for levels 1, 3, and 5 for each of the 5 core dimensions (Clarity, Simplification, Warmth, Patience, Fluency).

• Calibration: Large Language Models tend to default to a "nice" median (e.g., scoring everyone a 4/5). Grounding the prompt in concrete behaviors forces the LLM to evaluate actual evidence from the transcript rather than outputting "vibes-based" scores.
• Explainability: This ensures the AI can cite specific quotes as evidence to justify its verdict to the human recruiter.

5. Managing Token Limits

We used a strict TPM (Tokens Per Minute) budget, specifically designing the prompt to work within tight context windows.

• Rolling Window: For the conversational /api/chat endpoint, we only send the last 4 messages (the sliding window) plus the turn-aware system prompt. This keeps latency lightning-fast and API costs near zero, while still providing enough context for Priya to sound coherent and follow up on the immediate previous answer.
• Full Context for Assessment: The full transcript is only sent once, at the very end of the interview, to the /api/assess endpoint for the final 5-dimension evaluation.

6. What I Would Do Differently with More Time

• Server-Side Audio (WebSockets + WebRTC): The Web Speech API is heavily Chrome-dependent and struggles with thick accents compared to Whisper. I would migrate to a WebSocket-based streaming architecture using a custom STT model to ensure cross-browser compatibility (Firefox/Safari).
• Dynamic Turn Limits: Instead of a hard stop at 7 turns, allow Priya to end the interview early if she detects the candidate has definitively failed (e.g., using inappropriate language) or successfully demonstrated all competencies.
• Vector DB for RAG Calibration: Connect the assessment LLM to a vector database containing historical, human-graded interview transcripts to dynamically calibrate the AI's scoring strictness against the human recruiting team's real-world preferences.