latency-budget

A three-stage recommender under a 50 ms deadline: two-tower retrieval, HNSW search, LambdaMART reranking. Measured p99 of 20.5 ms at 50 req/s, sheds quality instead of timing out under load, catches a corrupted index in one canary cycle, corrects position bias in clicks. CPU-only, offline.

Ground PoC of ComputeFollowsPower (IDEA-001): dispatch latency-tolerant jobs to the cheapest-power node within their latency budget and verdict whether the power arbitrage beats the movement cost.

Real-time authorization decisions measured under injected failure: timeouts, caching, graceful degradation and a kill switch, measured rather than asserted. Whether a timeout bounds the tail, what fail-open versus fail-closed costs once both errors are priced, and whether webhook-driven decisioning is idempotent under replay.

Dialogue runtime for a voice agent: policy as a transition table, barge-in that never resumes, per-stage latency budget at p95, line capacity, speakable-reply guardrails. Dependency-free, no vendor calls.

Production-grade AI latency budgeting and reactive scaling framework for LLM inference systems. Covers p50/p95/p99 modeling, SLO design, Kubernetes (K8s) HPA patterns, and distributed AI infrastructure. By Vipin Kumar