ZEDX AI Simulator is a professional interview simulation and training platform engineered for rigorous candidate evaluation. It provides a comprehensive Dual-Mode system (Web & Desktop) integrating Speech-to-Text (STT), Large Language Model (LLM) inference, and a hybrid Text-to-Speech (TTS) engine.
- The Web Simulator (Rigorous Evaluation Mode): A strict, autonomous Voice-to-Voice AI Agent that conducts structured mock interviews in 30+ supported languages.
- The Desktop Sandbox (Assisted Learning Mode): A Windows-native environment where the AI acts as a real-time copilot to accelerate skill acquisition.
graph TD
subgraph CloudLayer ["Cloud Layer (Next.js Edge & Supabase)"]
AUTH["Opaque Auth Gateway"] --> DB[("Protected Storage with RLS")]
API["Edge API Serverless Routes"] --> CONTEXT_SERVER["Context Aggregation Node"]
CONTEXT_SERVER --> DB
end
subgraph LocalExecutionLayer ["Local Execution Layer (Electron OS Engine)"]
SYS["Local Microphone Audio Capture"] --> VAD["Voice Activity Detection Layer"]
MIC["Hardware Microphone Allocation"] --> VAD
IPC["IPC Secure Bridge"] --> API
end
subgraph NeuralInferenceLayer ["Neural Inference Layer (Groq LPU Cluster)"]
VAD -- "WebM Chunk Streaming" --> WHISPER["Whisper V3 Engine (Load Balanced x14)"]
WHISPER -- "Raw Parsed Transcripts" --> LLM["Qwen Engine (Load Balanced x14)"]
CONTEXT_SERVER -- "Full CV + JD Context" --> LLM
LLM -- "Actionable SSE Stream" --> IPC
end
subgraph TTSHybridLayer ["TTS Hybrid Router"]
LLM -- "AI Response Text" --> LANG_DETECT{"Language Tag Inspector"}
LANG_DETECT -- "ar-* prefix" --> ELEVEN["ElevenLabs Multilingual V2 (Load Balanced x10)"]
LANG_DETECT -- "All other languages" --> EDGE["Microsoft Edge TTS (Serverless Native, Free)"]
ELEVEN --> AUDIO_OUT["Browser Audio Context"]
EDGE --> AUDIO_OUT
end
classDef cloud fill:#0f172a,stroke:#3b82f6,stroke-width:2px,color:#fff;
classDef local fill:#1e293b,stroke:#10b981,stroke-width:2px,color:#fff;
classDef neural fill:#312e81,stroke:#6366f1,stroke-width:2px,color:#fff;
classDef tts fill:#3b0764,stroke:#a855f7,stroke-width:2px,color:#fff;
class AUTH,DB,API,CONTEXT_SERVER cloud;
class SYS,MIC,VAD,IPC local;
class WHISPER,LLM neural;
class LANG_DETECT,ELEVEN,EDGE,AUDIO_OUT tts;
flowchart TD
INPUT["AI Response Text + Language Tag"] --> ROUTER{"Language Router\n lang.startsWith('ar')"}
ROUTER -- "TRUE: Arabic Dialect" --> ELEVEN_POOL["ElevenLabs Key Pool\n[10 independent accounts]"]
ROUTER -- "FALSE: All Other Languages" --> EDGE_PROC["Edge TTS Engine\nMicrosoft Neural Voices\nZero quota cost"]
ELEVEN_POOL --> SHUFFLE["Fisher-Yates Shuffle\nPer-Request Randomization"]
SHUFFLE --> TRY_KEY["Attempt Request\nwith Key[i]"]
TRY_KEY -- "HTTP 200 OK" --> STREAM["Stream audio/mpeg\nto Browser AudioContext"]
TRY_KEY -- "HTTP 4xx/5xx" --> LOG["Log Key Failure\nAdvance to Key[i+1]"]
LOG --> TRY_KEY
TRY_KEY -- "All Keys Exhausted" --> ERROR["Return 500\nAll quota pools exhausted"]
EDGE_PROC --> TEMP_FILE["Write audio to\ntemp file (fs)"]
TEMP_FILE --> READ_BUFFER["Read Buffer\ninto Response"]
READ_BUFFER --> CLEANUP["Unlink temp files\n(finally block)"]
CLEANUP --> STREAM
classDef router fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px,color:#fff;
classDef eleven fill:#3b0764,stroke:#a855f7,stroke-width:2px,color:#fff;
classDef edge fill:#064e3b,stroke:#10b981,stroke-width:2px,color:#fff;
class ROUTER router;
class ELEVEN_POOL,SHUFFLE,TRY_KEY,LOG,ERROR eleven;
class EDGE_PROC,TEMP_FILE,READ_BUFFER,CLEANUP edge;
sequenceDiagram
participant REACT as React useEffect
participant REF as hasStartedRef
participant AUDIO as Global Audio Registry
participant TTS as TTS API Route
participant CTX as Browser AudioContext
REACT->>REF: Check hasStartedRef.current
alt Already initialized
REF-->>REACT: Return (suppress duplicate invocation)
else First invocation
REF->>REF: Set hasStartedRef.current = true
REACT->>AUDIO: Kill all active HTMLAudioElement instances
Note over AUDIO: Pause + src='' + remove from registry
REACT->>TTS: POST /api/tts {text, language, voice}
TTS-->>CTX: Stream audio/mpeg chunks
CTX->>CTX: Playback via HTMLAudioElement
CTX->>AUDIO: Register instance in global registry
end
Note over REACT,CTX: Any new speakText() call first\nexecutes global kill before streaming
sequenceDiagram
participant UI as React UI (Renderer)
participant OSE as Electron OS Capturer
participant VAD as VAD Middleware (Web Worker)
participant STT as Groq Whisper V3 (Rotated Key Pool)
UI->>OSE: Request Internal Audio Capture
OSE-->>UI: Grant MediaStream Track constraints
loop Continuous Buffering Loop
UI->>VAD: Stream compressed WebM chunks
alt Audio Intensity > Threshold Limit
VAD->>STT: Dispatch Blob Transduction Buffer
Note over STT: Key pool shuffle on every request
STT-->>UI: Return Parsed JSON Transcript Payload
else Absolute Silence Delta Reached
VAD-->>VAD: Purge internal buffer (Zero API Bloat)
end
end
graph LR
subgraph ContextAssemblyLine ["Context Assembly Line"]
HISTORY["Interview History & Memory\n(All prior Q&A)"]
JD["Job Description\n(Complete, untruncated)"]
RESUME["Candidate CV\n(Full document content)"]
ANGLE["Randomized Interview Angle\n(Behavioral/Technical/Situational)"]
NAME["Candidate Name\n(Extracted from CV)"]
end
subgraph PromptSynthesisEngine ["Prompt Synthesis Engine"]
COMPILER["Dynamic System Prompt Compiler\n+ Pronunciation Normalization"]
HISTORY --> COMPILER
JD --> COMPILER
RESUME --> COMPILER
ANGLE --> COMPILER
NAME --> COMPILER
end
subgraph ExecutionResolution ["Execution & Resolution"]
LLM["Qwen Engine\n(14-key Load Balanced Pool)"]
UI["Secure React Practice Interface"]
COMPILER -- "Full Contextual Semantic Prompt" --> LLM
LLM -- "Server-Sent Events (SSE)" --> UI
end
classDef data fill:#064e3b,stroke:#34d399,stroke-width:2px,color:#fff;
classDef engine fill:#701a75,stroke:#e879f9,stroke-width:2px,color:#fff;
class HISTORY,JD,RESUME,ANGLE,NAME data;
class COMPILER,LLM,UI engine;
- Zero Truncation Policy: The LLM operates on the complete, untruncated candidate data (CV + JD) to ensure maximum personalization.
- Pronunciation Normalization: System prompts dynamically adjust for Arabic vs Latin scripts to prevent TTS mispronunciations.
flowchart TD
US["User Boot Sequence"] -->|Opaque Publishable Key| NEXT["Next.js Hydration Context"]
NEXT -->|JWT Verification Cycle| SUPA["Supabase Engine"]
SUPA --> RLS{"Row-Level Security Evaluator"}
RLS -- Valid UUID ownership bounds --> ALLOW("Permit Access to User-Owned Records")
RLS -- Cryptographic Identity Mismatch --> DENY("Reject Unauthorized Access")
ALLOW --> DB[("Secure Transaction Logs & Transcripts")]
- AI Prompt Injection Guard: Encapsulation blocks malicious PDF resume instructions.
- Row-Level Security Vaulting: PostgreSQL policies mathematically bind data to verified users.
- API Key Isolation: All keys reside exclusively in server-side environments, using dynamic arrays for load balancing.
graph TD
subgraph PostInterviewAggregation ["Post-Interview Aggregation"]
STATE["Session State"] -->|Extract| TRANSCRIPT["Complete Spoken Transcript"]
STATE -->|Extract| METADATA["JD, CV, Duration, Language"]
TRANSCRIPT --> PAYLOAD["JSON Context Payload"]
METADATA --> PAYLOAD
end
subgraph EvaluationEngine ["Evaluation Engine (Groq LPU)"]
PAYLOAD -->|Next.js API Edge Route| GROQ["Groq Evaluator (Qwen)\n(Load Balanced Key Pool)"]
GROQ -->|Strict JSON Schema| ANALYSIS{"JSON Evaluator"}
ANALYSIS -- "Technical Score (0-10)" --> METRICS["Output State"]
ANALYSIS -- "Communication Score" --> METRICS
ANALYSIS -- "Ideal Benchmarks" --> METRICS
end
subgraph ClientRenderingExport ["Client Rendering & Export"]
METRICS -->|Hydrate| SCORECARD["React Scorecard UI"]
SCORECARD -->|Apply @media print styles| PRINT["Browser Print Spooler"]
PRINT -->|Strip UI & backgrounds| PDF["Organized PDF Report"]
end
classDef state fill:#451a03,stroke:#b45309,stroke-width:2px,color:#fff;
classDef ai fill:#312e81,stroke:#6366f1,stroke-width:2px,color:#fff;
classDef render fill:#064e3b,stroke:#10b981,stroke-width:2px,color:#fff;
class STATE,TRANSCRIPT,METADATA,PAYLOAD state;
class GROQ,ANALYSIS,METRICS ai;
class SCORECARD,PRINT,PDF render;
- Strict JSON Enforcement: Forces the evaluation LLM to output predictable schema-validated JSON, preventing parsing crashes in the UI.
- Client-Side Hydration & Export: Heavy lifting (rendering and PDF compilation) is offloaded entirely to the client's browser, saving server-side compute.
sequenceDiagram
participant USER as User (Microphone)
participant STT as Groq Whisper V3 (x14 Key Pool)
participant LLM as Qwen Engine (x14 Key Pool)
participant ROUTER as TTS Language Router
participant ELEVEN as ElevenLabs (x10 Key Pool, Arabic)
participant EDGE as Edge TTS (Free, All Others)
participant CTX as Browser AudioContext
USER->>STT: Speaks (VAD Filtered Audio Chunks)
Note over STT: Sub-second Transcription\nRotated key selected via shuffle
STT-->>LLM: Forward Transcribed Text
Note over LLM: Full CV + JD Context Injected\nPronunciation directives applied
LLM-->>ROUTER: Stream response text with language tag
ROUTER->>ROUTER: Inspect language tag prefix
alt Arabic (ar-*)
ROUTER->>ELEVEN: POST with shuffled key from pool
ELEVEN-->>CTX: Stream audio/mpeg
else All other languages
ROUTER->>EDGE: Execute Native Edge TTS
EDGE-->>CTX: Buffer audio response
end
CTX->>USER: Real-time Audio Playback\n(Global mute gate applied)
Committed to engineering strict software architectures that bridge the gap between autonomous ML deployments and instantaneous human accessibility.