---
schema: 1
type: video
title: >-
  Learned Execution Graphs for Anomaly Detection & Drift in APIs — Ritvik
  Pandya, JP Morgan Chase
description: >-
  Traditional monitoring reported the system healthy: latency down, errors at
  zero. A mandatory processing step had been silently skipped, and nothing
  caught it except the graph. Ritvik Pandya's team at JP Morgan models each API
  request as a short lived execution graph, a DAG of the middleware steps it
  passes through, learned from telemetry at over 1,600 requests per second.
  Compare what actually ran against that learned graph and a skipped, reordered,
  or injected step stops hiding behind healthy averages. The same graph
  localizes performance problems to the exact node instead of the whole
  endpoint. In production it flagged a 41x deviation at a single node that
  service level monitoring never saw, cutting root cause from hours to under 30
  seconds. The talk separates a one off anomaly from real drift, a slow shift
  that needs a new baseline, and sorts drift into structural, volume, and
  behavioral, using per node baselines and KL divergence rather than one
  threshold for every request. The payoff is a cheap tier one check that only
  escalates when the graph says something actually changed. Speaker info: -
  https://www.linkedin.com/in/ritvik-pandya/ Timestamps: 0:00 - Execution graphs
  for anomaly and drift detection 1:07 - What a short lived execution graph is
  3:28 - Tiered checks and per client baselines 5:23 - The method: baseline,
  deviation, localize, act 6:16 - Localizing a slow node, and how the system is
  trained 7:33 - Anomaly versus drift 8:55 - The three kinds of drift:
  structural, volume, covariate 12:46 - The pipeline: from telemetry to gradual
  rollout 13:54 - Hot path versus recon, and worked examples 15:21 - Tuning it:
  delayed events, sampling, cold starts 17:09 - Results and lessons
canonical: 'https://liquidgold.cc/watch/u1yaOeEX4e8'
markdown: 'https://liquidgold.cc/watch/u1yaOeEX4e8.md'
llms: 'https://liquidgold.cc/watch/u1yaOeEX4e8/llms.txt'
parent: 'https://liquidgold.cc/channel/aidotengineer.md'
channel: aidotengineer
updated: '2026-10-24T06:54:02.555Z'
language: null
children: []
videoId: u1yaOeEX4e8
durationSeconds: 1178
publishedAt: '2026-07-23T00:00:04.000Z'
searchChunks:
  - ordinal: 0
    text: null
    startSeconds: 81
    endSeconds: 214
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=81'
    heldReason: rights
  - ordinal: 1
    text: null
    startSeconds: 81
    endSeconds: 214
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=81'
    heldReason: rights
  - ordinal: 2
    text: null
    startSeconds: 214
    endSeconds: 227
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=214'
    heldReason: rights
  - ordinal: 3
    text: null
    startSeconds: 227
    endSeconds: 326
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=227'
    heldReason: rights
  - ordinal: 4
    text: null
    startSeconds: 227
    endSeconds: 326
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=227'
    heldReason: rights
  - ordinal: 5
    text: null
    startSeconds: 326
    endSeconds: 339
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=326'
    heldReason: rights
  - ordinal: 6
    text: null
    startSeconds: 339
    endSeconds: 384
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=339'
    heldReason: rights
  - ordinal: 7
    text: null
    startSeconds: 396
    endSeconds: 453
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=396'
    heldReason: rights
  - ordinal: 8
    text: null
    startSeconds: 453
    endSeconds: 464
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=453'
    heldReason: rights
  - ordinal: 9
    text: null
    startSeconds: 464
    endSeconds: 543
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=464'
    heldReason: rights
  - ordinal: 10
    text: null
    startSeconds: 543
    endSeconds: 554
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=543'
    heldReason: rights
  - ordinal: 11
    text: null
    startSeconds: 554
    endSeconds: 770
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=554'
    heldReason: rights
  - ordinal: 12
    text: null
    startSeconds: 554
    endSeconds: 770
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=554'
    heldReason: rights
  - ordinal: 13
    text: null
    startSeconds: 781
    endSeconds: 837
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=781'
    heldReason: rights
  - ordinal: 14
    text: null
    startSeconds: 781
    endSeconds: 837
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=781'
    heldReason: rights
  - ordinal: 15
    text: null
    startSeconds: 848
    endSeconds: 928
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=848'
    heldReason: rights
  - ordinal: 16
    text: null
    startSeconds: 848
    endSeconds: 928
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=848'
    heldReason: rights
  - ordinal: 17
    text: null
    startSeconds: 938
    endSeconds: 1029
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=938'
    heldReason: rights
  - ordinal: 18
    text: null
    startSeconds: 938
    endSeconds: 1029
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=938'
    heldReason: rights
  - ordinal: 19
    text: null
    startSeconds: 938
    endSeconds: 1029
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=938'
    heldReason: rights
  - ordinal: 20
    text: null
    startSeconds: 1029
    endSeconds: 1042
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=1029'
    heldReason: rights
  - ordinal: 21
    text: null
    startSeconds: 1042
    endSeconds: 1178
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=1042'
    heldReason: rights
  - ordinal: 22
    text: null
    startSeconds: 1042
    endSeconds: 1178
    watch: 'https://liquidgold.cc/watch/u1yaOeEX4e8?t=1042'
    heldReason: rights
---


# Learned Execution Graphs for Anomaly Detection & Drift in APIs — Ritvik Pandya, JP Morgan Chase

Traditional monitoring reported the system healthy: latency down, errors at zero. A mandatory processing step had been silently skipped, and nothing caught it except the graph. Ritvik Pandya's team at JP Morgan models each API request as a short lived execution graph, a DAG of the middleware steps it passes through, learned from telemetry at over 1,600 requests per second. Compare what actually ran against that learned graph and a skipped, reordered, or injected step stops hiding behind healthy averages. The same graph localizes performance problems to the exact node instead of the whole endpoint. In production it flagged a 41x deviation at a single node that service level monitoring never saw, cutting root cause from hours to under 30 seconds. The talk separates a one off anomaly from real drift, a slow shift that needs a new baseline, and sorts drift into structural, volume, and behavioral, using per node baselines and KL divergence rather than one threshold for every request. The payoff is a cheap tier one check that only escalates when the graph says something actually changed. Speaker info: - https://www.linkedin.com/in/ritvik-pandya/ Timestamps: 0:00 - Execution graphs for anomaly and drift detection 1:07 - What a short lived execution graph is 3:28 - Tiered checks and per client baselines 5:23 - The method: baseline, deviation, localize, act 6:16 - Localizing a slow node, and how the system is trained 7:33 - Anomaly versus drift 8:55 - The three kinds of drift: structural, volume, covariate 12:46 - The pipeline: from telemetry to gradual rollout 13:54 - Hot path versus recon, and worked examples 15:21 - Tuning it: delayed events, sampling, cold starts 17:09 - Results and lessons

Canonical: https://liquidgold.cc/watch/u1yaOeEX4e8 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md)

> Search chunk text is held until public-source provenance is verified.

## Document chunks

> Missing: no stored document chunks are available.