karvifi/foundation/ai-engineering

response = await client.messages.create(
    model="claude-sonnet-4-5",    # pinned model version
    max_tokens=2048,               # ALWAYS set — prevents runaway cost
    system=system_prompt,
    messages=[{"role": "user", "content": sanitized_input}],
    timeout=30.0,                  # ALWAYS set
)
# Validate output before using
result = OutputSchema.model_validate_json(response.content[0].text)

Agent = LLM + Tools + Memory + Goal + Stop Condition
 
Required:
- max_tool_calls: 20 (budget cap)
- max_iterations: 10 (loop cap)
- Full trace logging (Langfuse)

# Use cheapest model that works
classify → claude-haiku-3    # cheap
write_code → claude-sonnet-4-5  # mid-tier
architect → claude-opus (rarely)  # expensive

☐ max_tokens set on every LLM call
☐ model version pinned
☐ rate limiting on AI endpoints
☐ output validated with Pydantic/Zod
☐ user input sandboxed
☐ cost tracking enabled (Langfuse)
☐ eval test suite exists (DeepEval)
☐ prompts versioned in prompts/ directory