Chat with my CV!

Intelligent, Interactive AI Agent with Autonomous Lead Logging

Architecture · Challenges · Learnings

---

Outcome

Built an autonomous LLM agent with tool-use capabilities for professional representation. Implemented structured output schema enforcing JSON response contracts, designed multi-step agentic workflow (intent classification, entity extraction, validation, external API action), and grounded all responses in YAML knowledge base to eliminate hallucination.

Technical Build

Architected a context-injection approach where the full professional profile is embedded in the system prompt, ensuring 100% factual grounding with zero hallucination risk. Designed JSON response contract enabling deterministic routing between conversational replies and autonomous tool execution. Integrated Google Sheets API for real-time lead capture triggered by natural language intent detection.

---

The Core Mechanism

User Question → Claude API → JSON Response → Action Router
                    ↓
            Profile Context
            (System Prompt)

Response Type	Trigger	Action
reply	General questions	Return conversational response
log_lead	Hiring intent detected	Log to Google Sheets + confirm

---

Agentic Workflow

The autonomous multi-step task demonstrates agentic behavior:

flowchart LR
    subgraph Perception["Step 1: Perception"]
        P1["Parse user message"]
        P2["Detect intent signals"]
        P1 --> P2
    end

    subgraph Reasoning["Step 2: Reasoning"]
        R1["Classify intent"]
        R2["Select action type"]
        R1 --> R2
    end

    subgraph Action["Step 3: Action"]
        A1["Extract entities"]
        A2["Validate data"]
        A3["Call external API"]
        A1 --> A2 --> A3
    end

    subgraph Confirm["Step 4: Confirm"]
        C1["Generate response"]
        C2["Return to user"]
        C1 --> C2
    end

    Perception --> Reasoning --> Action --> Confirm

    style Perception fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    style Reasoning fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    style Action fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    style Confirm fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    linkStyle 0,1,2,3,4,5,6,7 stroke:#FFFFFF,stroke-width:2px

Loading

Intent Triggers:

• "Can we schedule a call?"
• "What's your email?"
• "Send me your resume"
• "We'd like to interview you"

---

Key AI Engineering Concepts

Concept	Implementation
Agentic AI	Autonomous agent with perception-action loop
Tool Use	External API calls (Google Sheets) triggered by LLM
Structured Outputs	JSON schema contract for deterministic behavior
Knowledge Grounding	YAML source-of-truth preventing hallucination
Multi-Step Reasoning	Intent, Extract, Validate, Act, Confirm
Guardrails	"Do not invent facts" instruction enforcement

---

Architecture

flowchart TB
    subgraph S1["User Interface"]
        A1["Streamlit Web App"]
        A2["Chat Interface"]
        A1 --> A2
    end

    subgraph S2["Agent Core"]
        B1["Claude API Integration"]
        B2["JSON Response Parser"]
        B1 --> B2
    end

    subgraph S3["Context Injection"]
        C1["profile.yaml"]
        C2["System Prompt Builder"]
        C1 --> C2
    end

    subgraph S4["Autonomous Actions"]
        D1["Intent Detection"]
        D2["Google Sheets Logger"]
        D1 --> D2
    end

    S1 --> S2
    S3 --> S2
    S2 --> S4

    style S1 fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    style S2 fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    style S3 fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    style S4 fill:#1a1a2e,stroke:#A78BFA,color:#A3B8CC
    linkStyle 0,1,2,3,4 stroke:#FFFFFF,stroke-width:2px

Loading

---

Structured Output Contract

The agent returns structured JSON for deterministic action routing:

# Normal reply - for answering questions
{"type": "reply", "message": "Your conversational response here"}

# Lead logging - when follow-up intent detected
{"type": "log_lead",
 "company": "company name or null",
 "contact_name": "their name or null",
 "contact_email": "their email or null",
 "role_title": "role discussed or null",
 "notes": "brief conversation summary"}

---

Project Structure

interactive-ai-agent/
├── app/
│   ├── app.py              # Streamlit web interface (HF entry point)
│   ├── agent.py            # Core chat agent with Claude API
│   ├── prompts.py          # System prompt templates
│   ├── tools.py            # Profile loading, lead logging
│   ├── profile.yaml        # Professional profile data
│   └── requirements.txt    # App-specific dependencies
├── requirements.txt        # Full dependencies (local dev)
├── .gitignore
├── LICENSE
└── README.md

---

Usage

# Install dependencies
pip install -r requirements.txt

# Set environment variables
export ANTHROPIC_API_KEY="your-api-key"
export GOOGLE_SHEETS_ID="your-sheet-id"  # Optional

# Run Streamlit app
streamlit run app/app.py

---

Methods & Techniques

Category	Methods
Prompt Engineering	Persona injection, role definition, constraint setting, output schema enforcement
Structured Outputs	Response contract, type discrimination, nullable fields, schema validation
Intent Detection	Keyword triggers, semantic understanding, context awareness
Entity Extraction	In-context extraction, structured output mapping, null handling
Knowledge Grounding	Single source of truth, prompt injection, negative instructions

---

Environment Variables

Variable	Required	Description
ANTHROPIC_API_KEY	Yes	Claude API key
GOOGLE_SHEETS_ID	No	Sheet ID for lead logging
GOOGLE_CREDENTIALS_FILE	No	Path to service account JSON

---

License

MIT

---

Part of the AI/ML Portfolio

Return to Home | LinkedIn