Posted inAI AWS Machine Learning
Day 85: Time-Series and Anomaly Copilots
Time-series anomaly copilots should call CloudWatch/Prometheus as tools — not paste dashboard screenshots into prompts.
Day 84: SQL and Warehouse Copilots
SQL and warehouse copilots must be read-only by default and run EXPLAIN before any heavy query.
Posted inAI
Day 83: Multimodal Coding: Screenshots of Broken UIs
Multimodal coding from broken-UI screenshots needs redaction and evidence bundles — never paste raw PII pixels into the model.
Posted inAI
Day 82: Voice and Realtime Agents
Voice and realtime agents need turn-taking, barge-in, and clear non-voice defaults — do not force mic UX on every workflow.
Posted inAI
Day 81: Long-Term Memory Without Becoming Creepy
Long-term memory needs consent, TTLs, and user-visible entries — otherwise your agent becomes creepy surveillance.
Posted inAI AWS System Design
Day 80: Project: Platform AI Gateway
Build a platform AI gateway: one internal API for auth, budget, routing, and traces that every squad must use.
Day 79: FinOps Dashboards Engineers Open
Ship FinOps dashboards engineers open: cost per merged PR and per squad, not a monthly invoice surprise.
Posted inAI AWS System Design
Day 78: Load Shedding When the Model Is Sick
When the model is sick, shed load with feature flags and useful fallbacks — never return a silent timeout or a half-tool mutation.
Day 77: Caching RAG Answers Safely
Cache identical RAG answers only within a tenant and corpus version — never reuse answers across tenants or after index drift.
Posted inAI AWS System Design
Day 76: p99 of Agents: Queueing, Not Just Model Latency
Agent p99 is mostly queueing and human-gate wait — instrument TTFB, tool wait, and approval wait separately before you blame the model.