Echo — AI Representative & Resume Persona

Echo is an autonomous, RAG-grounded AI representative that acts as a personal interviewer and representative.The system is available over both a web chat interface and a live voice/phone call interface, enabling recruiters to ask questions about skills, experience, and projects, and schedule real-time interviews directly into Google Calendar.

---

🌐 Live Touchpoints

• Web Chat Interface: Live Chat App
• Voice Representative: Call +1 (254) 261-0787

---

🏗️ System Architecture

graph TB
    subgraph User_Touchpoints [User Touchpoints]
        Phone[Phone Call]
        Browser[Chat Interface]
    end

    subgraph Voice_Pipeline [Voice Pipeline - Vapi]
        VapiOrch[Vapi Orchestrator]
        STT[Deepgram Nova-3]
        LLM_V[GPT-4o Mini]
        TTS[ElevenLabs Turbo]
        Phone --> VapiOrch
        VapiOrch --> STT --> LLM_V --> TTS
        TTS --> VapiOrch
    end

    subgraph Chat_Pipeline [Chat Pipeline - Vercel]
        ChatUI[Chat UI]
        ChatAPI[API /api/chat]
        LLM_C[GPT-4o Mini via OpenAI]
        Browser --> ChatUI --> ChatAPI --> LLM_C
    end

    subgraph Shared_Services [Shared Services]
        RAG[RAG Retriever]
        VDB[Pinecone Vector DB]
        Cal[Cal.com API]
        RAG --> VDB
        RAG --> Cal
    end

    LLM_V --> RAG
    LLM_V --> Cal
    ChatAPI --> RAG
    ChatAPI --> Cal

    subgraph Data_Ingestion [Data Ingestion]
        Resume[Resume PDF]
        GitHub[GitHub Repos]
        Embed[OpenAI Embeddings]
        Resume --> Embed
        GitHub --> Embed
        Embed --> VDB
    end

Loading

---

🛠️ Technical Tech Stack

Component	Choice	Details
Voice Platform	Vapi	Deepgram Nova-3 (STT) + ElevenLabs Turbo v2.5 (TTS)
Chat LLM	GPT-4o Mini	High reasoning speed, budget friendly, excellent tool calling
Embeddings	OpenAI text-embedding-3-large	3072-dimensional semantic vectors
Vector DB	Pinecone Serverless	Fast semantic RAG indexing and similarity query
Calendar Platform	Cal.com API v2	Direct webhook-driven scheduling integration
Frontend Framework	Next.js 15 (App Router)	Server-side rendering, streaming API routes, responsive UI
Deployment	Vercel (Hobby)	Serverless edge handlers, automatic GitHub deployments

---

📈 Evaluation Summary (1-Page Report)

A professional 1-page evaluation report is available in the workspace: 📄 Evaluation Report PDF

Key Benchmarks

• First-Response Latency: Mean: 1.22s | P95: 1.85s (Well under the 2.0s limit)
• Chat Groundedness: 95.5% (Standard 33 Qs) | 94.1% (Rigorous 38 Qs stress-test)
• Chat Accuracy: 93.6% (Standard 33 Qs) | 94.1% (Rigorous 38 Qs stress-test)
• Keyword Accuracy: 100% (Correctly answers identity, skills, and projects in all 38 test cases)
• Transcription Accuracy: 99.1% (Deepgram Nova-3)
• Booking Completion Rate: 100% (6/6 test bookings confirmed on Cal.com)
• Hallucination Rate: 0.0% (Zero cases flagged by GPT-4o Judge in both runs)
• Security Leak Rate: 0.0% (Jailbreak and prompt injection attempts blocked)

Operational Cost Metrics

• Voice Call: ~$0.123 / call (STT + TTS + LLM + Telephony)
• Chat Session: ~$0.004 / session (Pinecone Vector retrieval + LLM tokens)

---

🚀 Setup & Installation

1. Prerequisites

Ensure you have Node.js 18+ and npm installed.

2. Environment Setup

Copy the example environment file and populate the secrets:

cp .env.example .env.local

Fill in the following variables:

• OPENAI_API_KEY
• PINECONE_API_KEY
• CALCOM_API_KEY
• CALCOM_EVENT_TYPE_ID
• VAPI_SECRET_KEY

3. Ingest Personal & GitHub Data

Run the ingestion pipeline to parse the resume PDF, crawl GitHub repositories, embed the content, and upload to Pinecone:

npm run ingest:all

4. Running the Development Server

Start the local server:

npm run dev

5. Running Evaluations

Generate the chat and voice evaluation data:

npm run eval:chat
npm run eval:voice

Compile the final PDF/Markdown evaluation report:

npm run eval:report

Note: Run node evals/eval-report.js to compile the report directly without compilation overhead.

---

🛡️ Production Security & Fallbacks

• Local Context Fallback: If Pinecone lookup fails or times out, the backend gracefully falls back to a pre-built local JSON copy of key resume facts, preventing agent silence.
• Cal.com API Fallback: If booking fails programmatically, the representative immediately provides a direct public calendar link.
• Input Filtering: Incoming chat messages are filtered against prompt injection and system leakage attacks.
• Webhook Authorization: Webhook endpoints are guarded by secret-token validation header x-vapi-secret to reject unauthorized requests.