pipeline trace visualization

Data through the Atom

Follow one observation from raw Mac metrics through quantization, token encoding, transformer attention, anomaly scoring, to final verdict. Six stages. One pipe. The complete KIRI pipeline in motion.

Press Step or Auto Play to begin
stage 1
OBSERVE
CPU
--
Memory
--
Disk
--
Swap
--
Load
--
Network
--
stage 2
QUANTIZE
stage 3
ENCODE
stage 4
ATTEND
stage 5
SCORE
stage 6
VERDICT
The complete pipeline in atom.py
# The complete pipeline in ~10 lines
obs = collect_local()           # OBSERVE
tokens = lang.encode(obs)       # QUANTIZE + ENCODE
for tok in tokens:              # ATTEND
    logits = atom.forward(tok, pos, keys, vals)
prob = atom._softmax(logits)    # SCORE
score = -log(prob[actual])      # VERDICT
KIRI — Intelligent Runtime Inspector
Built on Karpathy's microgpt · Open Source · 2026