Real-Time Transformer Prediction Pipeline
At Autodesk, a high-throughput transformer ran on a real-time event pipeline that analyzed more than one million daily user interactions for next-click prediction.
- Verified outcome
- 1M+ daily interactions
Confidentiality note. Employer-specific implementation details, internal names, and proprietary scientific context have been intentionally generalized. The scale, responsibilities, technologies, and outcomes below are limited to facts already published in Sudhanva’s résumé and professional profile.
Problem
A next-click prediction workload needed model inference in the path of a high-volume event stream rather than as an offline research exercise.
Scale and constraints
Scale
- More than one million user interactions analyzed each day.
- Real-time event processing with transformer inference.
Constraints
- Keep the prediction path compatible with a real-time event pipeline.
- Balance transformer throughput with the operational needs of an existing product system.
Sudhanva's ownership
As a Machine Learning Engineer Intern, Sudhanva deployed the high-throughput transformer into the event pipeline.
- Transformer serving work.
- Integration with the real-time prediction pipeline.
- Throughput-oriented deployment.
Architecture at a safe level
The public architecture is intentionally described through system responsibilities rather than internal service names or proprietary topology. The relevant boundaries were workload execution, orchestration, data movement, observability, and delivery.
- Treated serving and event integration as part of the model-delivery problem.
- Focused optimization on throughput at the observed daily interaction volume.
Important engineering decisions and tradeoffs
- Real-time serving places tighter latency and failure constraints on model execution than offline analysis.
- Only publicly documented system boundaries are included here; product-specific architecture remains private.
Results
- Deployed the transformer on the production event path.
- Supported analysis of more than one million daily interactions.
Reliability and operations
- The workload operated as part of a real-time event pipeline rather than a manual batch process.
Transferable lessons
- Model serving begins with the latency, throughput, and failure contract of the product path.
- Production integration work is as important as model selection for real-time ML.
Technologies
- Transformers
- Real-time event processing
- Model serving