Cloud & GenAI Platform
 at Optic 2000

Context

During my alternance at Audioptic Trade Services (ATS), the technology subsidiary of the Optic 2000 group, I worked within the Digital Factory team on two major pillars: building Serverless B2B applications for opticians and prototyping an AI chatbot that went into production.

The group operates 1,400+ optical stores across France. The Digital Factory team’s challenge was to deliver scalable internal tooling without a dedicated ops team — hence the bet on AWS Serverless.

LLM Chatbot — Atsia

The flagship project: Atsia, a production chatbot built on Amazon Bedrock (Anthropic Claude model) to assist opticians with frequently asked operational questions — reducing inbound support calls to the ATS helpdesk.

Architecture

Optician (browser)
       │
       │ HTTPS
       ▼
┌──────────────────────────────────────────────────────────┐
│               AWS API Gateway (REST)                     │
└─────────────────────────┬────────────────────────────────┘
                          │ Lambda proxy integration
                          ▼
┌──────────────────────────────────────────────────────────┐
│               AWS Lambda (Java / Spring Boot)            │
│  - Prompt engineering + context injection                │
│  - Conversation history management (DynamoDB)            │
│  - Bedrock InvokeModel call                              │
└──────┬──────────────────┬───────────────────────────────┘
       │                  │
       ▼                  ▼
Amazon Bedrock       DynamoDB
(Claude Sonnet)      (session history)

Bedrock Integration (Java)

// BedrockService.java
@Service
public class BedrockService {

    private final BedrockRuntimeClient bedrockClient;
    private final PromptBuilder promptBuilder;

    public ChatResponse invoke(ChatRequest request, List<Message> history) {
        String prompt = promptBuilder.buildWithContext(request.message(), history);

        InvokeModelRequest invokeRequest = InvokeModelRequest.builder()
            .modelId("anthropic.claude-3-sonnet-20240229-v1:0")
            .body(SdkBytes.fromUtf8String(buildAnthropicPayload(prompt)))
            .contentType("application/json")
            .build();

        InvokeModelResponse response = bedrockClient.invokeModel(invokeRequest);
        return parseClaudeResponse(response.body().asUtf8String());
    }

    private String buildAnthropicPayload(String prompt) {
        return """
            {
              "anthropic_version": "bedrock-2023-05-31",
              "max_tokens": 1024,
              "system": "%s",
              "messages": [{"role": "user", "content": "%s"}]
            }
            """.formatted(SYSTEM_PROMPT, escapeJson(prompt));
    }
}

Prompt Engineering — System Prompt

The system prompt is the core of Atsia’s quality. It was iteratively refined with the support team to ground the model in ATS-specific knowledge:

You are Atsia, the AI assistant for ATS (Audioptic Trade Services), 
the digital subsidiary of the Optic 2000 group.

Your role is to assist opticians with operational questions about:
- The B2B platform (orders, invoicing, stock)
- Product catalog and commercial conditions
- Technical issues with the portal

Rules:
- Always respond in French
- If you are not confident (< 80% certainty), say so explicitly and 
  redirect to the support hotline: 01 XX XX XX XX
- NEVER invent product references or prices
- Keep answers under 150 words unless a technical procedure requires more

Serverless Full-Stack Applications

In parallel, I built and maintained B2B applications for the optician network:

AWS Architecture Pattern

React/Redux SPA (S3 + CloudFront)
         │
         │ API calls
         ▼
API Gateway → Lambda (Java/Spring Boot) → RDS (MySQL) / S3
                                        → QuickSight (BI dashboards)

Lambda Cold Start Optimization

A key challenge with Java Lambdas is cold start latency. Solution: Provisioned Concurrency + Spring Boot optimization:

// LambdaHandler.java — SnapStart-aware initialization
@SpringBootApplication
public class LambdaApplication implements RequestHandler<APIGatewayProxyRequestEvent, APIGatewayProxyResponseEvent> {

    // Static initialization happens during SnapStart snapshot
    private static final ApplicationContext context =
        SpringApplication.run(LambdaApplication.class);

    @Override
    public APIGatewayProxyResponseEvent handleRequest(
            APIGatewayProxyRequestEvent event, Context ctx) {
        HttpHandler handler = AwsProxyHandlerBuilder
            .httpApiV2ProxyBuilder()
            .defaultProxy()
            .buildAndInitialize();
        // ...
    }
}

Using AWS Lambda SnapStart reduced cold start from ~3s to ~200ms for Java 21 runtimes.

Quality — Cypress E2E Tests

// cypress/e2e/chatbot.cy.js
describe('Atsia Chatbot', () => {
  beforeEach(() => {
    cy.loginAsOptician()
    cy.visit('/chatbot')
  })

  it('responds to a stock inquiry', () => {
    cy.get('[data-testid="chat-input"]')
      .type('Comment consulter mon stock disponible ?')

    cy.get('[data-testid="send-button"]').click()

    cy.get('[data-testid="bot-message"]', { timeout: 10000 })
      .should('be.visible')
      .and('contain.text', 'stock')

    // Ensure no hallucinated product references
    cy.get('[data-testid="bot-message"]')
      .invoke('text')
      .should('not.match', /REF-\d{6}/)
  })
})

Highlights at AWS Summit Paris 2025

The team’s work on Atsia was showcased internally at the AWS Summit Paris 2025. We benchmarked our architecture against approaches from Orange Bank (Ippon Technologies) and Veolia (12+ LLM models, 250+ knowledge bases, 60,000 users) — validating our technical choices and identifying scaling roadmap priorities.

Impact