"Know what you know. Know what breaks you."
Interview Mirror is an adaptive, metacognitive technical interview and skill profiling system. Traditional mock interviewers ask static question lists and rate candidates on superficial keywords. Interview Mirror does something fundamentally different:
- Calculates your Confidence Calibration Gap: Detects whether you are genuinely knowledgeable or dangerously overconfident.
- Predicts your Cognitive Breakpoint: Locates the exact architectural threshold where your reasoning degrades.
- Executes Adversarial Battle Drills: Puts you through high-pressure stress scenarios to force recovery and measure net resilience.
- Maintains Cross-Session Breethe Memory: Tracks your skill trajectory, flags stale technical knowledge, and personalizes every subsequent round.
flowchart TD
subgraph Client["Frontend Studio (React 19 + Vite + Three.js)"]
UI["Cyberpunk Studio UI"]
Radar["Dynamic SVG Skill Radar"]
GraphVis["LangGraph 6-Node Live Visualizer"]
Canvas["Three.js Living Network Canvas"]
end
subgraph Backend["Backend Engine (Express ESM + Node.js)"]
Router["API Gateway / Router"]
subgraph GraphEngine["LangGraph State Engine"]
N1["1. Evaluate Node"] --> N2["2. Calibrate Node"]
N2 --> N3["3. Belief State Engine"]
N3 --> N4["4. Metacognitive Profiler"]
N4 --> N5["5. Question Planner"]
N5 --> N6["6. Persona Critique"]
end
subgraph Services["Core Micro-Services"]
CalEngine["Calibration Engine"]
BayesEngine["Dirichlet / Beta Belief Engine"]
StyleEngine["Thinking Style Detector"]
FrustAgent["Anti-Cheat Socratic Escalator"]
end
end
subgraph Cloud["External AI & Cloud Memory"]
Nvidia["NVIDIA NIM\n(meta/llama-3.3-70b-instruct)"]
Breethe["Breethe Cloud Memory API\n(thebreeth.com)"]
SQLite["SQLite Embedded Store\n(better-sqlite3)"]
end
UI -->|REST / JSON| Router
Router --> GraphEngine
GraphEngine --> Services
Services --> Nvidia
Services --> Breethe
Services --> SQLite
GraphVis -.->|Node Telemetry| GraphEngine
The core interview loop is managed by a stateful 6-node LangGraph:
- Node 1 (Evaluate): Grades technical accuracy (0.0β1.0) and isolates missing architectural primitives.
-
Node 2 (Calibrate): Computes the mathematical delta (
$\Delta = \text{Confidence} - \text{Demonstrated Depth}$ ). - Node 3 (Update Belief): Bayesian probability update across 8 technical domains.
- Node 4 (Thinking Style Profiler): Classifies reasoning into First-Principles, Pattern Matcher, or Framework Applier.
- Node 5 (Planner): Synthesizes curriculum modules, stale topics from memory, and past weaknesses.
- Node 6 (Critique): Enforces interviewer voice and strictly validates question difficulty.
- Alex (System Design & Scalability): Socratic, warm, rigorous. Focuses on concurrency, data partitioning, and failure modes.
- Priya (ML & RAG Architecture): Direct, precise, first-principles. Drills vector search latency, evaluation metrics, and model tradeoffs.
- Marcus (Operational Resilience & MCP): Pragmatic, no-nonsense lead engineer. Focuses on edge cases, outages, and agent tools.
- 5-turn pressure drill targeting identified failure modes.
- Quantifies net recovery with a Before/After scorecard (e.g.,
$41% \rightarrow 57%, +16%$ recovery).
- Interactive algorithm & design challenges with progressive hints (
$0 \rightarrow 3$ ). - Never leaks direct solutions; enforces structured reasoning and self-reflection.
- Cross-session growth curves tracking readiness and calibration convergence.
- Stale topic decay detection automatically prompts review of unpracticed domains.
- Full semantic episode search via vector memory API.
- Zero-PII privacy audit with SHA-256 session anonymization.
| Layer | Technologies |
|---|---|
| Frontend | React 19, Vite 8, Framer Motion, Three.js, Lucide Icons, Vanilla CSS Glassmorphism |
| Backend | Node.js (ESM), Express, LangGraph (@langchain/langgraph), @langchain/core |
| Persistence | SQLite (better-sqlite3), Local JSON Disk Fallback |
| LLM Inference | NVIDIA NIM (meta/llama-3.3-70b-instruct), Gemini, OpenAI, Anthropic |
| Episodic Memory | Breethe Cloud Memory REST API (https://api.thebreeth.com/v1) |
| Testing | Node.js Native Test Runner (node --test), 42 Automated Integration Tests |
- Node.js: v18.0.0 or higher
- npm: v9.0.0 or higher
git clone https://github.com/Creater-jod/interview-agent.git
cd interview-agent
# Install backend dependencies
npm install
# Install frontend dependencies
npm install --prefix clientCreate .env in the project root (see .env.example for details):
PORT=5000
NODE_ENV=development
CORS_ORIGIN=*
# Primary LLM: NVIDIA NIM
LLM_PROVIDER=nvidia
LLM_MODEL=meta/llama-3.3-70b-instruct
NVIDIA_API_KEY=your_nvidia_nim_api_key_here
# Breethe Memory (optional, falls back automatically to local persistent JSON)
BREETHE_API_KEY=your_breethe_api_key_here
BREETHE_API_BASE=https://api.thebreeth.com/v1For frontend customization, create client/.env:
VITE_API_URL=http://localhost:5000# Terminal 1: Backend Intelligence Engine (:5000)
npm run dev
# Terminal 2: Frontend Studio (:5173)
npm run dev --prefix clientOpen http://localhost:5173 in your browser.
The repository includes a comprehensive 42-test automated suite covering controllers, LangGraph pipelines, calibration mathematics, Breethe memory, and privacy audits:
npm testExpected output:
# tests 42
# suites 0
# pass 42
# fail 0
A pre-configured render.yaml blueprint is included in the root directory:
- Connect your GitHub repository to Render.
- Create a new Web Service from the repo root.
- Configure environment variables (
NVIDIA_API_KEY,BREETHE_API_KEY). - Build command:
npm install - Start command:
node src/index.js - Health check path:
/api/health
A pre-configured vercel.json is included for zero-config SPA routing:
- Import the repository into Vercel.
- Set Root Directory to
client(or use root with default build script). - Add environment variable:
VITE_API_URL=https://your-backend-service.onrender.com - Deploy!
| Method | Endpoint | Description |
|---|---|---|
GET |
/api/health |
Healthcheck & service status |
POST |
/api/interview/start |
Initialize interview session with chosen persona & track |
POST |
/api/interview/continue |
Submit response + confidence; runs 6-node LangGraph turn |
POST |
/api/interview/complete |
Conclude session and generate Diagnostic Dossier |
GET |
/api/interview/:id/messages |
Fetch complete transcript from SQLite store |
GET |
/api/report/:sessionId |
Retrieve readiness, calibration delta, and breakpoint card |
POST |
/api/battle/turn |
Execute pressure turn in Adversarial Battle drill |
GET |
/api/memory |
Retrieve candidate growth trajectory, stale topics, and traits |
GET |
/api/memory/status |
Check Breethe Cloud sync vs Local Disk fallback status |
POST |
/api/memory/search |
Semantic vector search over candidate past episodes |
POST |
/api/train/start |
Launch adaptive Socratic coaching session |
POST |
/api/train/message |
Submit message or solution proposal to Socratic Coach |
POST |
/api/train/hint |
Request progressive hint tier escalation (0β3) |
POST |
/api/train/new-problem |
Request next adaptive challenge matched to profile |
GET |
/api/train/profile |
Fetch persistent skill profile and radar data |
This project is licensed under the MIT License β see the LICENSE file for details.