From ed2195d87962c69ec5519834d29100c26ebbe762 Mon Sep 17 00:00:00 2001
From: Viktor Barzin <me@viktorbarzin.me>
Date: Fri, 22 May 2026 20:25:19 +0000
Subject: [PATCH] feat(meet-kevin): throttle inter-video LLM calls (30s) to
 stay under Anthropic RPM

First production run hit Anthropic's per-account rate_limit_error (429) trying
to burn through 16 backfill videos in seconds. The SDK's built-in retry can't
recover because the rate limit window resets slower than the 3 retry attempts.

Added meet_kevin_inter_video_sleep_seconds (default 30s) to PipelineDeps and
main's _process_pending_videos loop. 16 backfill videos now take ~8 min (16 * 30s
sleeps + ~30s per LLM call) instead of bursting into the rate limit.
---
 services/meet_kevin_watcher/config.py   | 1 +
 services/meet_kevin_watcher/main.py     | 6 +++++-
 services/meet_kevin_watcher/pipeline.py | 3 +++
 3 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/services/meet_kevin_watcher/config.py b/services/meet_kevin_watcher/config.py
index 99d55e9..b5e8d53 100644
--- a/services/meet_kevin_watcher/config.py
+++ b/services/meet_kevin_watcher/config.py
@@ -21,6 +21,7 @@ class MeetKevinWatcherConfig(BaseConfig):
     meet_kevin_llm_model: str = "claude-sonnet-4-5"
     meet_kevin_prompt_version: str = "v1"
     meet_kevin_daily_cost_cap_usd: float = 5.0
+    meet_kevin_inter_video_sleep_seconds: int = 30
 
     # API credentials
     anthropic_oauth_token: str = ""
diff --git a/services/meet_kevin_watcher/main.py b/services/meet_kevin_watcher/main.py
index 0e8f96a..79b167f 100644
--- a/services/meet_kevin_watcher/main.py
+++ b/services/meet_kevin_watcher/main.py
@@ -104,7 +104,10 @@ async def _process_pending_videos(
         )
         videos = result.scalars().all()
 
-    for video in videos:
+    for i, video in enumerate(videos):
+        # Throttle between videos to avoid bursting into Anthropic's per-account RPM limit
+        if i > 0:
+            await asyncio.sleep(deps.inter_video_sleep_seconds)
         async with session_factory() as session:
             # Re-fetch inside its own session so we can commit per-video
             result = await session.execute(
@@ -207,6 +210,7 @@ async def run() -> None:
         prompt_version=config.meet_kevin_prompt_version,
         daily_cost_cap_usd=Decimal(str(config.meet_kevin_daily_cost_cap_usd)),
         workdir=config.meet_kevin_workdir,
+        inter_video_sleep_seconds=config.meet_kevin_inter_video_sleep_seconds,
     )
 
     # Graceful shutdown
diff --git a/services/meet_kevin_watcher/pipeline.py b/services/meet_kevin_watcher/pipeline.py
index 0f114ef..93fe6e1 100644
--- a/services/meet_kevin_watcher/pipeline.py
+++ b/services/meet_kevin_watcher/pipeline.py
@@ -66,6 +66,9 @@ class PipelineDeps:
     workdir: str
     """Filesystem directory for yt-dlp caption downloads."""
 
+    inter_video_sleep_seconds: int = 30
+    """Sleep between consecutive videos to stay under the LLM provider RPM limit."""
+
 
 # ---------------------------------------------------------------------------
 # Daily cost accounting