kapavamoda.buzz • OpenAPI 3.1

Execute inference — routed, admission-controlled, watermarked.

Optimized routing with enterprise-grade reliability. You pick the target model, we ensure consistent performance. Monitor your latency in queue, prefill and decode metrics.

Quickstart — 3 steps before 200 OK

Operator provisions x-kapa-key scoped to cohort-03.
Export KAPA_KEY and KAPA_BASE=https://kapavamoda.buzz
POST with routing hints — see runnable example →

TTSR target <5min. Sandbox vs prod isolated. Limited cohort keeps goodput high.

Auth — Bearer + x-kapa-key

All calls require Authorization: Bearer $KAPA_KEY plus cohort header x-kapa-cohort: 03. Keys are scoped, rotated by operator, audit-logged.

Scope

inference:execute

Rate

per-key RPS + TPM

Residency

KV-cache 1M

Concepts — why TTFT vs TPOT matters

Prefill is compute-bound (prompt ingestion, MLA compression). Decode is memory-bound (KV-cache streaming). TTFT is your prefill SLO. TPOT is your decode slope. Batch max_num_seqs too high and you trade RPS for queuing.

TTFT p50 186msTPOT 14.2msGoodput 99.2%

Operator note: MoE 1.04T total / 32B active - you pay the 32B inference cost, not the 1T marketing number. Same for DeepSeek 671B/37B, Qwen 235B/22B.