NOC · ShortFlix prod · asia-northeast3 Day 14 / 30 SLO: 99.91% (target 99.5%) Burn rate: 1.7× On-call: jisu (sole)
GCP $500 credit
$ 386.10 / 500
▲ $24.40 / day · runs out D+18
RapidAPI bill
$ 132.40
ceiling $200/mo · 66% used
p50 latency
1.18 s
budget 1.5 s · 21% headroom
p95 latency
1.94 s
budget 2.0 s · 3% headroom
Cost / DAU / mo
$ 0.038
ceiling $0.05 · 24% margin
Open incidents
1
SEV-3 · trend-safety latency

$500 GCP credit ledger · 30-day projection

$ 386.10 / $ 500.00 · D+14 of 30 · projected EOM $ 632 (over by $132)
Ceiling guard: if projected EOM > $500, auto-cut Vertex grounding to ko-only and downgrade trend-safety to flash-lite. Lever owner: jisu.
LineSo farDaily run-rateProjected EOMCeiling
Cloud Run · asia-NE3 (min=1)$ 88.20$ 6.30 / d$ 189$ 220
Gemini-2.0-flash$ 121.40$ 8.67 / d$ 260$ 240
Vertex AI Search$ 41.10$ 2.94 / d$ 88$ 100
Cloud SQL f1-micro$ 4.20$ 0.30 / d$ 9$ 12
Egress · CDN$ 18.40$ 1.31 / d$ 39$ 60
Logging · Cloud Tasks · misc$ 12.80$ 0.91 / d$ 27$ 40
GCP subtotal$ 286.10$ 20.43 / d$ 612$ 500 credit
3rd-party (out of credit)So farProjected EOMCeiling
RapidAPI · YT Shorts plan$ 49.00$ 49.00$ 60
RapidAPI · IG Reels plan$ 49.00$ 49.00$ 60
RapidAPI · TikTok plan$ 34.40$ 88.00$ 80
RapidAPI subtotal$ 132.40$ 186.00$ 200

SLO burn · 7-day window

orchestrator p95budget 2.0 s
1.94 s
curator agent p95nightly batch
3m12s
trend-safety p95budget 600 ms
714 ms
unified-search p953 MCP fanout
312 ms
RapidAPI 5xx rateover all 3 plans
2.1 %

Open alerts

!
SEV-3 · trend-safety p95 over budget
14:08 KST · firing 22 min · runbook: switch flash-lite
+19%
SEV-4 · TikTok RapidAPI 5xx spike
11:46 KST · 16 occurrences · 3 retries succeeded
2.1%
RESOLVED · curator nightly batch overran 60s window
04:22 KST · increased Cloud Scheduler timeout 60→120s
closed
RESOLVED · Vertex AI Search 429
02:14 KST · backed off 2x · within budget
closed

Today's runbook

# 14:30 KST · operator: jisu $ sf ops cost --ledger → projected EOM exceeds $500 by $132 — engaging tier-1 lever $ sf ops lever apply tier1 --tag credit-ceiling + vertex.search corpus = ko-only (was ko+en+ja) + trend-safety = gemini-2.0-flash-lite (was -flash) ! novelty drift expected: 0.86 → ~0.83 (acceptable, > 0.80 floor) $ sf ops slo refresh → p95 1.94s · burn 1.7× → 0.9× by 16:00 $ sf ops lever rollback --in 24h scheduled rollback at 14:30 +1d if EOM projection back under $500

Multi-agent superiority on the ops axis: each lever (corpus, model size, fanout depth) maps to one agent — single-agent has no levers, only a kill switch.

Capacity plan · D+14 → D+30

RiskTriggerLeverOwner
$500 credit overrunEOM > $500Tier 1 (corpus + flash-lite)jisu
RapidAPI ceilingTikTok plan > $80nightly cache only on TTjisu
p95 over budgetp95 > 2.0 s 5 minparallel fanout disablejisu
Solo-dev fatigueon-call paged 3+ in 24hSEV-3 → email-onlyjisu
Demo day freezeD+27 onwardcode freeze · CFR-only deployjisu

No headcount. No PagerDuty. One operator, one Cloud Run service, four agents, six levers, $500 ceiling.