A field guide for Solution Architects & Product Owners
Architectural Traits of AI
Treat every model as a black box with four traits: its shape (what goes in, what comes out), its cost, its latency, and its behaviour with constraints. This guide organizes those traits into three pillars — and into the workflow you'll actually run every time a requirement lands on your desk.
The workflow: from requirement to production decision
Each step of the workflow is one pillar of the guide — this page is the map, the pillars are the tools.
What you do when a requirement lands
The three pillars
Each pillar has its own detail page with decision diagrams, traits and cheat sheets.
1 · Model Capabilities & Selection
Don't learn how models work inside — learn their input → output matchings. Requirement says "detect defects in factory video"? You should instantly think CNN object detection (YOLO), not GAN.
- The major model families and what each is shaped for
- Choose-it-when / avoid-it-when traits per family
- A requirement → model cheat sheet
2 · Model Operational Metrics
The math you do need — operational, not algorithmic. Context windows, latency vs throughput, VRAM sizing (~2 GB per billion parameters at FP16), and how to read a Data Scientist's evaluation report.
- Context window limits and token budgets
- Latency vs throughput trade-offs
- GPU memory & cost rules of thumb
- Precision, recall, F1, FID — judging production readiness
3 · Integration Patterns & Paradigms
How models plug into enterprise systems: RAG to connect LLMs to your data securely, and the cost ladder of prompt engineering → RAG → fine-tuning.
- RAG architecture, benefits and failure modes
- When prompting is enough vs when to fine-tune
- Cost / effort comparison of the three approaches
The questions you'll actually be asked
Stakeholder question → your move → where the answer lives.
| When they ask… | The architect's move | Where |
|---|---|---|
| "Can AI solve this?" | Translate the requirement into an I/O shape, then match it to a family | Model selection — decision diagram |
| "Which model should we shortlist?" | Pick the variant inside the family — with named models per type | Family deep dives |
| "What will it cost to run?" | VRAM & token arithmetic: params × bytes, tokens × price, headroom | Operational metrics — VRAM & context |
| "Why is the pilot so slow?" | Separate latency from throughput; check TTFT and workload mixing | Operational metrics — latency |
| "Can it use our internal data — securely?" | RAG with access control enforced at the retrieval step | Integration — RAG architecture |
| "Should we fine-tune?" | Walk the escalation ladder: prompt → RAG → fine-tune, only on proven failure | Integration — the ladder |
| "DS reports 97% accuracy — can we ship?" | Ask which error is expensive, then check that metric, not the average | Operational metrics — evaluation |
Quick orientation: which family for which job
The 10-second version of pillar 1 — full traits and examples on the detail page.
| Business requirement | Reach for | Typical models |
|---|---|---|
| Understand / generate text, answer questions | Transformer LLM (+ RAG for your data) | GPT-4, Claude, Llama |
| Detect objects or defects in images / video | CNN-based vision model | YOLO, ResNet |
| Generate images from text | Diffusion model | Stable Diffusion, DALL·E |
| Synthetic data, upscaling, face generation | GAN | StyleGAN, ESRGAN |
| Predict from business tables (churn, fraud, credit) | Classic ML — gradient boosting | XGBoost, LightGBM |
| Semantic search, similarity, clustering | Embedding model + vector DB | text-embedding models, CLIP |