ABTalks AI Cohort Interview Agent

An adaptive, intelligent technical interview platform designed to evaluate cohort candidates dynamically based on their progress through the 31-day AI engineering curriculum.

The platform features a premium dark-themed user interface, dynamic difficulty scaling, custom visual neural branding, and a comprehensive automated testing suite.

---

System Architecture

The Interview Agent bridges candidates' curriculum progress, real-time response evaluations, and adaptive next-step recommendations through a stateless and serverless-compatible design:

graph TD
    subgraph Client [React Frontend]
        UI[Magic Patterns Dark UI]
        SessionHook[useInterviewSession]
        Report[Feedback Report Dashboard]
    end

    subgraph Server [TypeScript Backend]
        API[Express Router]
        Store[Session Store]
        Planner[Interview Planner]
        Evaluator[Answer Evaluator]
        Generator[Question & Follow-up Generator]
        LLM[Gemini API / Local Keyword Fallback]
    end

    UI -->|Start Session / Submit Answer| SessionHook
    SessionHook -->|POST /api/interview| API
    
    API -->|Restore State from Client| Store
    API -->|Evaluate Response| Evaluator
    API -->|Plan Curriculum Path| Planner
    
    Evaluator -->|Substance & Keyword Check| LLM
    Generator -->|Difficulty-Adaptive Prompts| LLM
    
    API -->|Return State & Payload| UI
    Store -->|Dynamic Feedback & Next Steps| Report

Loading

---

Core Features

1. Adaptive Interviewing Engine

• Dynamic Difficulty Alignment: Interviews scale and calibrate question depth (Foundational, Intermediate, Advanced) based on candidate profile signals and answer correctness.
• Curriculum-Aware Routing: Ensures candidate evaluations cover at least 4 unique curriculum days and cross a minimum of 8 comprehensive questions.
• Context Preservation: Maintains complete conversation state to ask logical, cohesive follow-up questions without repeating previously explored topics.

2. Answer Evaluation & Feedback Engine

• Substance Over Length: Evaluates technical precision and concrete concepts (such as specific tool usage, scaling parameters, or memory limits) rather than simple response word count.
• Question-Aware Keyword Analysis: In rate-limit or API fallback scenarios, responses are graded against specific technical keywords customized to the question context (such as failover retry logic, Prometheus monitoring metrics, or vector database indexing architectures).
• Granular Performance Metrics: Generates scores across 5 primary dimensions: Technical Understanding, Problem Solving, Communication, Depth, and Practical Application.

3. Premium Enterprise UI & Brand Identity

• Custom Geometric Neural Branding: Features the BrandIcon component—an abstract neural head outline forming a speech bubble, housing a Y-branched connection map. Used consistently across the navbar, chat turn markers, and the pulsating AI evaluation indicator.
• Dynamic Learning Links: The "Review Topic" button on candidate feedback report cards routes the candidate directly to the Learning Progress dashboard, automatically highlighting the cohort module requiring reinforcement.

4. Serverless Session Persistence

• Stateless Flow: Designed to operate reliably in serverless environments (such as Vercel Serverless Functions) where backend instances are ephemeral.
• Client-Side State Tracking: The React client stores the latest session configuration state (sessionState) and passes it along with each successive /api/interview POST request.
• Backend State Restoration: The backend dynamically restores the session context prior to executing any evaluation or question-generation logic, ensuring complete state stability without requiring persistent database connections.

---

Getting Started

1. Environment Setup

Create a .env file in the project root:

PORT=5000
GEMINI_API_KEY=your_gemini_api_key_here
GEMINI_MODEL=gemini-2.5-flash

2. Install Dependencies

npm install

3. Start Development Server

Launches both the backend server (port 5000) and frontend bundler (Vite):

npm run dev

4. Build Production Assets

Compiles the frontend assets to the dist/ directory and compiles the serverless backend function bundle:

npm run build

---

Testing Suite

The project includes 67 automated test cases checking backend state rules, dynamic evaluation engines, and edge-case handling.

Run tests using:

npm test

Coverage Highlights:

• planner.test.ts: Verifies dynamic cohort pathing, adaptive difficulty alignment, and candidate personalization rules.
• feedbackEngine.test.ts: Verifies qualitative performance summaries, strengths/gaps mapping, and dynamic next-steps selection.
• fabricatedFeedbackBug.test.ts: Ensures empty responses are blocked, incomplete sessions compile partial evaluations correctly, and session states do not leak between candidate switches.
• concisenessPolish.test.ts: Enforces word count thresholds, prevents chatty preambles, and runs semantic duplicate checks on generated questions.