THEPerformance Machine.

Perfloop learns your system from code and telemetry, and hunts a wide catalog of performance patterns. A change reaches you only after your checks pass, the speedup is measured against the baseline, and disproving it has failed. What you get is a stream of proven pull requests, and software that stays fast.

watch the loop run · 3:58 →

the controls · four switches per repository · one research cadence

perfloop.ai·/t/oss/frontier

Frontier

Cases
HypothesisWherePatternState
Quadratic repository matching scan in selectRepoSet for branch-targeted queries
perfloop/zoekt
not anchored
INEFFICIENT ALGORITHMVERIFIED
Return early for constant sorting keys
perfloop/clickhouse
MergeTree insert part-write loop
INEFFICIENT ALGORITHMVERIFIED
Pool OTLP converter scratch state
perfloop/prometheus
OTLP metrics ingestion
ALLOCATION HOT LOOPVERIFIED
BCE, loop unrolling, and redundant copy elimination in lazyIORBitmap and lazyORBitmap
perfloop/roaring
Parallel union operations (OR)
DATA PARALLEL GAPVERIFIED
Eliminate cache-populate completion wait after upstream EOF
perfloop/tag
S3 object GET cache delivery
SYNC WAITVERIFIED

proof

merged upstream

running perfloop

  • Tigris Data

model: 129 repos333 cases run · 75 validated across 40 repos · 23 merged upstreaminvestment: 3,286 measured patch revisionssnapshot · 2026-08-24

Start the loop.
Keep it fast.