AI 工程该建什么：Jev 式 System-One 决策路由排第一
一份 AI 工程构建清单，首推用 Jev 式类型化决策加置信度升级搭建自己的 System-One 决策路由。
For people who keep asking what to build in AI Engineering. Part 2. The meta moved in 3 weeks. Here's what matters now. ➣ Build your own System-One Decision Router (Jev-style typed decisions + confidence escalation to System-2) ➣ Build your own KV-Cache-Aware Gateway (prefix caching + session-affinity routing + TTFT tracking) ➣ Build your own Speculative Decoding Pipeline (local draft model + cloud verify + acceptance-rate telemetry) ➣ Build your own Inter-Agent Sanitization Proxy (cross-agent injection defense + per-agent permission scoping) ➣ Build your own Shadow Routing Comparator (5% traffic mirror + trajectory and cost diffs) ➣ Build your own Computer-Use Fallback Agent (API-first, headless browser when no API exists) ➣ Build your own PII-Safe Inference Proxy (UUID masking before context + unmasking after + audit logs) ➣ Build your own Cost Kill-Switch Middleware (projected-spend halts + frontier → mid → local → cache degradation) ➣ Build your own Real-Time Voice Agent Loop (sub-1s interruption handling + streaming tool calls) ➣ Build your own DPO Preference Pipeline (thumbs-down → pairs → nightly LoRA → eval-gated promotion) ➣ Build your own Memory Eviction Engine (3-tier memory + importance scoring + TTL policies) ➣ Build your own Trajectory Replay Debugger (record every hop + time-travel debugging for agents) ➣ Build your own HITL Approval Gateway (checkpointed approvals via Slack/Teams + resume tokens) ➣ Build your own Agent Chaos Suite (fault injection: timeouts, malformed tools, context overflow) ➣ Build your own On-Device Inference Shim (WebGPU/MLX local fallback when the cloud goes down) Pick 3. Build them from scratch. Document every decision. Bookmark this. The meta won't wait.
这条是引用着另一条写的：x.com/suraj_sharma14/status/2097301427957948778