Concentric System Topology (AI ³ ML ³ DL ³ GenAI ³ Agents)
Active: Machine Learning
Latency Budget
15 ms
Est. Cost / 1M Reqs
$0.30
Execution Modality
Tabular
Primary Primitive
XGBoost / LightGBM
Warning: Using an LLM or Agent for low-latency tabular classification introduces a 50x latency penalty ($0.30 vs $25.00/1M ops) without accuracy improvement.
Custom Workload Evaluator
Engineered Rules
Input Data Modality
Tabular (Relational / Logs)
Latency SLA Constraint
< 20ms (Real-time in-line)
Reasoning & Tool Calling Loop
No (Static Evaluation)
Architectural Tier Trade-Off Matrix
| Layer | Core Definition | Representative Stack | Compute & Cost | Failure Mode & Defense |
|---|---|---|---|---|
| AI (Broad) | Rule-based expert systems, heuristic search, automation envelopes. | State machines, A* Search, Drools, Knowledge graphs | Minimal CPU ($0.01 / 1M) | Fragile to edge cases. Fix: Fallbacks & human triage. |
| Machine Learning | Statistical optimization & tabular pattern extraction from historical samples. | XGBoost, Scikit-learn, Random Forests, LightGBM | Low CPU/GPU ($0.25 - $1.00 / 1M) | Data drift, class imbalance. Fix: Retraining pipelines. |
| Deep Learning | Hierarchical neural representations for perception, audio & dense embeddings. | PyTorch, TensorRT, ResNet, ViT, Whisper | Dedicated GPU ($5.00 - $15.00 / 1M) | Distribution shift, adversarial noise. Fix: Robust data augmentation. |
| Generative AI / LLMs | Autoregressive foundation transformers synthesizing novel unstructured tokens. | Llama 3, Claude 3.5, GPT-4o, vLLM, SGLang | High H100 clusters ($50 - $400 / 1M) | Hallucinations, non-determinism. Fix: Grounded RAG, structured outputs. |
| AI Agents | Dynamic goal-directed loops with state, memory, tool orchestration & reflection. | LangGraph, CrewAI, AutoGen, MCP servers | Very High ($200 - $1500 / 1M) | Infinite loops, cascading errors. Fix: Execution budgets, sandboxing. |
📄 Generated Production Architecture Decision Record (ADR-0027)
Loading ADR content...