AI Gateway1 / 5
Can you replay
an AI call
as a DAG?
Or is observability still
“how many tokens did I spend”?
“how many tokens did I spend”?
apinizer.Every AI call measured and tracedAI Gateway2 / 5
An AI request is
not one call.
Failovers
Tool calls
Streaming chunks — it’s a chain
apinizer.Every AI call measured and tracedAI Gateway3 / 5
Latency,
broken down right.
01
TTFT
Time to first token
02
TPOT
Time per output token
03
p95 total
Which one is “slow” — see them apart
apinizer.Every AI call measured and tracedAI Gateway4 / 5
Spend anomaly,
before the invoice.
Not a static threshold — moving average + Bollinger band.
+3.2σ spend spike
→ runaway agent loop
→ budget capped
→ team notified
→ runaway agent loop
→ budget capped
→ team notified
apinizer.Every AI call measured and tracedAI Gateway5 / 5
Counting tokens is accounting,
not observability.
If you can’t see why it slowed, where the money went, which step broke — you’re not operating it, just paying for it.
Can you see step by step
why an AI call slowed down?
why an AI call slowed down?
apinizer.Every AI call measured and traced← → to navigate