Ajay Kumar Pondugala

Flagship · Runnable Lab

Distributed Scale Validation Lab

Separate Go submission API, transactional-outbox dispatcher, and RabbitMQ workers. PostgreSQL records accepted jobs, bounded attempts, terminal status, and one durable result per job.

Real-service tests cover known helper-process exits, broker restart, duplicates, and database rollback. Corrected local measurements retain every trial and include observation/polling overhead; this is a single-host synthetic lab.

Across twelve 1,000-job trials with 1/2/4/8 independent workers: 262.73–323.21 jobs/s and 95.43–151.08 ms observed p95. All measured jobs completed. Polling is approximately 50 ms plus query/sampling overhead; the run does not demonstrate linear scaling.

The historical 10,000-job result is an in-memory functional run, separate from the external-path evidence.

Corrected local evidence ↗

Watch the 90-second recovery demo ↗

The video is a captioned replay of actual agent-executed terminal output, not learner narration. Its 5.119233-second Compose/broker recovery is a separate trial from the corrected benchmark receipt.

Recovery evidence · Runnable Lab

Fault-Tolerant Transformer Lab

Four actual SIGKILL drills on pinned public data restore the exact six-step model, optimizer, RNG, and history state of an independent CPU control.

The interactive replay explains a committed report. It is a tiny local CPU proof at known failure boundaries, with no distributed-training or power-loss claim.