Portfolio | Samagra D. | AI Agent Architect

Architecture

Interactive Case Study: View implementation source code on GitHub

        graph TD
    %% Define Nodes with Click Events
    User((User Email)) --> Gmail[Gmail API]
    Gmail --> Supabase[(Supabase DB)]
    Supabase --> Router{LLM Router}
    
    %% Routing logic
    Router --> Gemini[Gemini 1.5]
    Router --> Claude[Claude 3.5]
    Router --> Groq[Groq Llama 3]
    
    Gemini --> Validator[Pydantic Validator]
    Claude --> Validator
    Groq --> Validator
    
    Validator --> Drafter[Drafting Agent]
    Drafter --> CrewAI[CrewAI Orchestration]

    %% Apply Click Events to specific nodes
    click Router call showDetails("Router") "View Router Logic"
    click Validator call showDetails("Validator") "View Validator Logic"
    click Drafter call showDetails("Drafter") "View Drafter Logic"

Trade-offs

Discussion on the trade-offs made in the architecture...

Core Components & Reasoning

LLM Router: The entry point. Reasoning: Abstracting the provider allows for seamless switching without impacting the core business logic.
Pydantic Validator: Used to enforce output structure. Benefit: Ensures downstream tools receive clean, typed data, reducing runtime crashes.
CrewAI Orchestration: Orchestrates the "Validator" and "Drafter" agents. Trade-off: Increased latency vs. higher quality control.

Skills Demonstrated

✅ LLM API Integration (Gemini, Claude, Groq)
✅ Graceful Fallback Design (error handling)
✅ Cost Optimization (Groq free-tier strategy)
✅ Latency & Rate Limit Strategy

Cost Analysis

Breakdown of expenses and RLS policy efficiency...

Model	Cost/1K	Speed	Reliability
Gemini Flash	Free	Fast	⚠️ Low
Claude Haiku	$0.15	Med	✅ High
Groq Llama	Free	Ultra-Fast	✅ High

Insight: Groq provides the best balance of speed and cost, making it the primary choice for high-volume processing.

Failure Scenarios

Handling strategies for LLM timeouts and errors...

Gemini Rate Limit (429): The router detected the limit at 10 AM, immediately switched to Claude, resulting in zero user-facing downtime.
Claude Credit Depletion: When $5 free credits were exhausted, the system automatically routed to the Groq free tier, keeping the process live.
Groq Model Deprecation: After a model was deprecated mid-project, I tested and migrated to llama-3.1-8b-instant, demonstrating the value of abstracting LLM calls.