Day 74: Batch Inference Overnight

Chat APIs are the wrong hammer for repo-wide refactors. Use Bedrock batch inference (or equivalent overnight jobs) to process large file sets cheaply, then apply intents in controlled daytime merges.

⚡ TL;DR: Queue file shards → batch model job → intents JSONL → daytime apply with CI. Not for interactive chat.

Batch shape

# batch/shard.py
def shards(paths: list[str], n=50):
    for i in range(0, len(paths), n):
        yield paths[i:i+n]

{"recordId": "f-001", "modelInput": {"messages": [{"role": "user", "content": "Emit rename intents for…"}]}}

# submit Bedrock batch (illustrative)
aws bedrock create-model-invocation-job \
  --job-name refactor-2026-09-11 \
  --role-arn arn:aws:iam::123:role/bedrock-batch \
  --model-id anthropic.claude-… \
  --input-data-config s3InputDataConfig={s3Uri=s3://batch/in/} \
  --output-data-config s3OutputDataConfig={s3Uri=s3://batch/out/}

Daytime apply

Never auto-merge overnight batch output. Humans + CI apply intents in PRs sliced by directory.

Closing checklist

• [ ] Shard inputs
• [ ] Batch for offline refactors
• [ ] Validate intents before apply
• [ ] Cost-compare vs interactive
• [ ] Keep human merge gates

Series navigation

Day 73: Draft-Then-Verify Routing · Day 75: Autoscaling Retrieval and Embed Jobs

Last updated September 11, 2026