Serving, Infra & Guardrails¶
The operational layer: ensuring the system doesn't bankrupt the company, leak data, or go down at 3am. Where applied scientists become legitimate systems engineers.
Rapid Recall
Seven cost levers, ranked by impact: model routing (3-5x), prompt caching (up to 90% on cached prefixes), output cap + structured outputs, semantic caching, context trimming, batch API (50% off), distillation. vLLM is the 2026 default serving engine; PagedAttention + continuous batching = 2-23x throughput. FP8 is the modern quantization default on H100+. Four serving modes: sync (one user), async (concurrent), streaming (TTFT matters), background (long jobs). Observability = traces with per-span input/output/cost; alert on p95 latency, error rate, cost-per-query, CSAT, hallucination rate. Guardrails are defense in depth: input (PII, injection, topic, rate), output (faithfulness, toxicity, PII redaction, schema), execution (step limits, cost limits, HITL, confidence escalation).
The production stack at a glance¶
flowchart LR
R[Request] --> IG[Input guardrails<br/>injection / PII / topic / rate]
IG --> RT[Router<br/>cheap vs frontier model]
RT --> CC[Semantic cache check]
CC -- hit --> RESP[Cached response]
CC -- miss --> SE[Serving engine<br/>vLLM or API]
SE --> OG[Output guardrails<br/>faithfulness / toxicity / PII / schema]
OG --> RESP
SE --> TR[Trace store<br/>Langfuse / LangSmith]
OG --> TR
TR --> AL[Alerts on drift]
Section guide¶
| Page | Covers |
|---|---|
| Cost & Token Economics | Token economics, the seven levers, model routing, quantization, cost dashboard |
| Inference Stack | vLLM, PagedAttention, continuous batching, parallelism strategies, alternative engines, hardware cost |
| Serving Modes | Sync, async, streaming, background jobs; state stores and durable execution |
| Observability | Tracing, alerts, feedback loops, shadow traffic |
| Guardrails | Input, output, execution guardrails; jailbreak defenses; privacy and compliance |
Layer Checklist¶
- Can you list seven cost-reduction levers ranked by impact?
- Can you explain PagedAttention and continuous batching in 90 seconds each?
- Can you configure vLLM for a production deployment?
- Can you calculate break-even volume for self-hosting vs API?
- Can you compare FP16, FP8, INT8, INT4 quantization tradeoffs?
- Can you design comprehensive guardrails (input, output, execution)?
- Can you defend against 3 specific prompt injection attack patterns?
- Can you design a feedback loop from user thumbs-down to eval set?
- Can you investigate a cost spike systematically?
- Can you design multi-provider failover for provider outages?