J.P. Morgan Uses Learned Execution Graphs to Catch Hidden API Failures in Production
AI Engineer · youtube · 2026-07-23
Learned execution graphs can catch skipped or reordered API steps
Ritvik Pandya from J.P. Morgan described a production monitoring approach that models each API request as a short-lived execution graph — a DAG of middleware steps inferred from telemetry at more than 1,600 requests per second.
- The system compares what actually ran against the learned baseline, so a skipped, reordered, or injected step can be detected even when latency and error-rate dashboards look healthy.
- In production, it found a 41× deviation at a single node that ordinary service-level monitoring missed, cutting root-cause analysis from hours to under 30 seconds.
- The talk distinguishes anomalies from drift, and breaks drift into structural, volume, and behavioral types.
- It uses per-node baselines and KL divergence instead of one global threshold, making the first-pass check cheap and only escalating when the graph shows a real change.
More from Infra
- A viral post says cheap public AI can now solve 80-year-old math problems — jacob_posel · 2026-07-23
- Enterprise AI is turning into a token factory, and utilization is the bottleneck — 新智元 · 2026-07-23
- Prime Intellect publishes 365,000+ unified RL tasks for SWE, terminal, and search agents — willccbb · 2026-07-23
- WSJ says AI chip startup Etched is in talks at a $20 billion valuation — Genzinvestor16180339 · 2026-07-23
- Can mixed AMD and Nvidia GPUs be pooled for inference at home? — Ecstatic-Wash-7667 · 2026-07-23
- Claude is now being used to scaffold a RISC-V CPU, SoC, and board workflow — globecsysinc · 2026-07-23