Minime: AI Personal Assistant

gfonseca.io

---

Minime is an AI-powered chat assistant that lives on my personal website. It acts as an AI version of me — able to talk about my background, projects, and even schedule meetings through natural conversation.

It combines:

• Large Language Models (LLMs) for dialogue
• Retrieval-Augmented Generation (RAG) from my Notion knowledge base
• A serverless, event-driven AWS architecture
• Infrastructure as Code (Terraform)
• CI/CD, monitoring, and security best practices

Features

• Conversational AI: Streamed responses via WebSocket for a natural chat feel
• Personalized knowledge: RAG retrieval from Notion content via Pinecone
• Meeting scheduling: Google Calendar integration for booking calls
• Geolocation awareness: Enriches user profile from IP location
• Email summaries: Sends conversation summaries via SES
• Rate limiting: DynamoDB-backed token bucket to protect resources (💰)
• Event-driven design: AWS EventBridge + SQS orchestration

System Architecture

Flow:

1. Frontend:
   • Static site on S3 served via CloudFront
   • Connects to backend via API Gateway (WebSocket) for low-latency streaming
2. Chat Handling:
   • chat_handler Lambda manages incoming messages, streams model output back
   • Rate limiting and error handling built-in
3. Retrieval:
   • Queries Pinecone with embeddings from Notion documents
4. LLM Invocation:
   • Uses Amazon Bedrock & Anthropic Claude
5. Backend Events:
   • EventBridge + SQS for async tasks like IP geolocation & conversation summaries
   • Summaries emailed via SES
   • Meetings scheduled in Google Calendar

Realtime Streaming Flow

sequenceDiagram
    autonumber
    participant U as User (Browser)
    participant CF as CloudFront + S3
    participant WS as API Gateway (WebSocket)
    participant CH as Lambda chat_handler
    participant C as chat.py
    participant VS as Pinecone (RAG)
    participant LLM as LLM Provider<br/>(Anthropic/Bedrock)
    participant GCal as Google Calendar
    participant DDB as DynamoDB (Conversations)

    U->>CF: Load app (HTML/JS/CSS)
    U->>WS: $connect
    WS->>CH: Connect event
    CH->>EventBridge: ConversationStart (ip)
    CH-->>WS: 200 Connected

    U->>WS: send {message}
    WS->>CH: route:message
    CH->>C: chat(session_id, message, on_stream)

    Note over C: Build memory window<br/>& running summary
    C->>DDB: get_conversation(), get_summary()
    C-->>DDB: messages, summary

    Note over C: RAG
    C->>VS: search(query)
    VS-->>C: top_k snippets

    Note over C: Build LLM messages (system+history+docs)
    C->>LLM: stream(messages, tools=[schedule_meeting, update_user_info])

    loop token stream
      LLM-->>C: text chunk
      C-->>CH: on_stream(chunk)
      CH-->>WS: {"op":"message_chunk", "content":chunk}
      WS-->>U: append to assistant bubble
    end

    alt tool call (schedule_meeting)
      LLM-->>C: tool_use(start_time, duration, summary, email)
      C->>GCal: create event
      GCal-->>C: htmlLink
      C->>LLM: tool_result(htmlLink)
      LLM-->>C: confirmation text (stream)
    end

    C->>DDB: add_message(user), add_message(assistant)
    C->>DDB: save_summary(updated)
    C-->>CH: return full text
    CH-->>WS: {"op":"finish"}
    WS-->>U: mark message done

Loading

Async Tasks: IP Geo & Summaries

sequenceDiagram
    autonumber
    participant WS as API Gateway (WebSocket)
    participant CH as Lambda chat_handler
    participant EB as EventBridge
    participant QG as SQS (geo)
    participant QSUM as SQS (summary)
    participant GEO as Lambda ip_geolocation_handler
    participant SUM as Lambda conversation_summary_handler
    participant GEOAPI as ip-api.com
    participant DDB as DynamoDB
    participant SES as Amazon SES

    WS->>CH: $connect
    CH->>EB: PutEvent ConversationStart {session_id, ip}
    EB->>QG: Rule -> SQS (geo)

    QG->>GEO: Invoke with {session_id, ip}
    GEO->>GEOAPI: Lookup(ip)
    GEOAPI-->>GEO: city, country, tz, zip
    GEO->>DDB: save_user_info(session_id, details)

    Note over CH: ...user chats...

    WS->>CH: $disconnect
    CH->>EB: PutEvent ConversationEnded {session_id}
    EB->>QSUM: Rule -> SQS (summary)

    QSUM->>SUM: Invoke with {session_id}
    SUM->>DDB: get_conversation(), get_summary(), get_user_info()
    SUM->>LLM: Summarise messages (short)
    SUM->>SES: send_email(from, to, "Conversation Summary", body)

Loading