This project modernizes a legacy interview platform with AI-assisted screening that reduces interviewer effort, improves role alignment, and delivers structured, auditable feedback — while keeping human review in the loop.
The system uses an AI Agent Orchestrator pattern with 5 specialized agents (Question Generator, Adaptive Follow-up, Response Evaluator, Feedback Synthesizer, Integrity Monitor) to automate the interview lifecycle from setup to scoring to review.
- USER_STORIES.md: 15 User Stories with acceptance criteria covering all 6 must-have features, prioritized P0/P1/P2.
- ARCHITECTURE.md: Target architecture diagram with Application Layer, AI/LLM Layer, Data Layer, and integration points. Technology stack decisions with rationale.
- SOLUTION_DESIGN.md: Key components, flows, assumptions, and 6 Architectural Decision Records (ADRs).
- PROMPTING_STRATEGY.md: 5 AI agents with full prompt specifications — intent, technique (Few-Shot, CoT, Rubric-Grounded), context window, expected output, and token budget.
- SEQUENCE_DIAGRAMS.md: 4 Sequence Diagrams — Interview Setup, Adaptive Session, Scoring & Feedback, Human Review.
- CLASS_DIAGRAM.md: 5 Class Diagrams — Core Domain, Scoring/Review, AI Agents, Audit/Cost, Data Store.
- ER_DIAGRAM.md: Full ERD with 9 tables, column specs, constraints, indexes, and data volume estimates.
- STATE_DIAGRAM.md: 5 State Machines — Interview, Question, Review Decision, Adaptive Questioning, Cost Tracking.
- ACTIVITY_DIAGRAM.md: 3 Activity Diagrams — Complete Interview Lifecycle, Question Generation Workflow, Cost Tracking Workflow.
- API_SPECIFICATION.md: RESTful API definitions for 6 service areas with JSON request/response examples and standard error format.
- TEST_STRATEGY.md: 50+ test cases mapped to user stories, AI-output validation strategy, 12 edge cases, automation vs manual matrix, security tests.
- GUIDING_PRINCIPLES.md: SOLID, KISS, YAGNI applied with domain-specific examples, anti-patterns avoided, and decision framework.
- interview-app/: Next.js web application demonstrating all 6 must-have features with premium dark-mode UI.
- AI Agent Orchestrator Pattern: Central coordinator for 5 specialized AI agents, each with its own prompt template, token budget, and single responsibility.
- Adaptive Questioning Engine: Chain-of-Thought reasoning to dynamically adjust question depth and topic coverage based on candidate responses.
- Human-in-the-Loop Design: AI generates recommendations with confidence scores; reviewers can Approve, Adjust, or Reject with full audit trail.
- Full Observability: Every AI prompt, response, and decision is logged with timestamps, token counts, and costs for complete traceability.
- Cost-Conscious GenAI: ~$0.0022 per interview (Gemini 2.0 Flash pricing); real-time token tracking with budget alerts.
- Modular Monolith for MVP: Clean module boundaries (services, agents, stores) designed for easy extraction to microservices post-MVP.
| # | Feature | Status |
|---|---|---|
| 1 | Flexible Interview Setup (role, skills, JD, resume) | ✅ Implemented |
| 2 | Adaptive Questioning (dynamic follow-ups based on depth) | ✅ Implemented |
| 3 | Automated Scoring & Feedback (structured rubric + narrative) | ✅ Implemented |
| 4 | AI Usage & Cost Tracking (tokens, costs per interview) | ✅ Implemented |
| 5 | Audit Trail & Usefulness Tracking (full AI decision log) | ✅ Implemented |
| 6 | Human-in-the-Loop Review (confidence scores + override) | ✅ Implemented |
| # | Feature | Status |
|---|---|---|
| 1 | Anti-cheating / Integrity Flagging | 📋 Documented |
| 2 | Dashboard for interview history and scores | ✅ Implemented |
| Layer | Technology |
|---|---|
| Frontend | Next.js 15 (App Router) |
| Styling | Vanilla CSS (Glassmorphism Dark Mode) |
| AI/LLM | Mocked AI Engine (Gemini-compatible) |
| Data | In-Memory Store (mirrors PostgreSQL schema) |
| State | React Context + useReducer |
cd interview-app
npm install
npm run devOpen http://localhost:3000 in your browser.
- Version: 1.0
- Status: MVP Complete
- Last Updated: June 2026