Summary
Ritvik Pandya, who leads the payments team at JP Morgan Chase, presented at AI Engineer on a production system built around what he calls learned execution graphs — short-lived directed acyclic graphs (DAGs) that model the complete request processing flow through a distributed system in real time. Unlike the persistent property graphs used by database products like Neo4j, these execution graphs are transient structures built per-request to enable anomaly detection and drift analysis across complex service meshes.
The architecture is conceptually straightforward: establish a baseline for how requests normally flow through layers including edge ingestion, authentication, orchestration, and parallel downstream services, then detect when request behavior deviates from that baseline. A tiered checking system minimizes overhead — requests that match the baseline pass through a lightweight first-tier check, while deviations trigger progressively deeper analysis to localize exactly which service node is the source of degradation. This allows the system to generate precise, actionable alerts rather than noisy volume-based thresholds.
Pandya categorizes three types of drift the system handles: structural drift (new nodes added or removed from the execution path), scale deviation (volume-driven latency changes requiring threshold recalibration), and covariate drift (shifts in request distribution over time, such as a growing proportion of international currency transactions). Benchmarking used OpenTelemetry and StarBench, injecting anomalies into millions of traces over a seven-day window to train the detection system ahead of live deployment. The approach is particularly well-suited to financial services workloads where pinpointing latency root causes across a distributed payments infrastructure is critical.
📺 Source: AI Engineer · Published July 23, 2026
🏷️ Format: Workflow Case Study
