Back office ops · Production

LinkedIn uses AI agents to accelerate model experimentation and TensorFlow-to-PyTorch migration

The problem

LinkedIn needed to migrate a large fleet of TensorFlow models to PyTorch and accelerate AI model iteration speed, while GPU compute was not expanding automatically to meet growing demand.

Workflow diagram · grounded in source
1
Agent receives task, parallelizes trials
trigger
“agents can parallelize model trials with minimal human input on an interactive dev machine”
2
Agent generates code iteratively
ai_action
“It runs an interative generate → score → hint → regenerate loop”
3
Verifier scores against quality gates
validation
“Each iteration is validated against explicit, rigorous quality gates, and the verifier doesn't just say 'fail'—it returns concrete, actionable fixes”
4
Structured feedback guides next generation
feedback_loop
“the verifier produces structured natural-language feedback that acts like an evaluation rubric plus a coach. Its job is to tell the agent not just that the generation is weak, but exactly where it is weak and what should be fixed next”
5
Scale via distributed training
integration
“Once the right model architecture is achieved, it is scaled via distributed training in an outer loop”
6
Promote to production via Flyte
output
“the PyTorch implementation is immediately validated on development GPU pods and promoted to production via Flyte workflows”
Reported outcome

LinkedIn's Autopilot for Torch agent drove higher productivity with much less manual effort across migration and auto-tuning workflows, and auto-tuning squeezed out 10%+ training throughput from already optimized LLM workloads.

Reported metrics
LLM training throughput improvement10%+
Manual effort for model migrationmuch less manual effort
Open-source benchmark performanceperformed strongly across open-source benchmarks, including more than 100 OpenML tasks
Reported stack
Autopilot for TorchTensorFlowPyTorchFlyte
Source
https://www.linkedin.com/blog/engineering/ai/ai-helping-build-better-ai-how-agents-accelerate-model-experimentation
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

LinkedIn's Autopilot for Torch agent drove higher productivity with much less manual effort across migration and auto-tuning workflows, and auto-tuning squeezed out 10%+ training throughput from already optimized LLM…

What tools did this team use?

Autopilot for Torch, TensorFlow, PyTorch, Flyte.

What results were reported?

LLM training throughput improvement: 10%+; Manual effort for model migration: much less manual effort; Open-source benchmark performance: performed strongly across open-source benchmarks, including more than 100 OpenML tasks (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Agent receives task, parallelizes trials → Agent generates code iteratively → Verifier scores against quality gates → Structured feedback guides next generation → Scale via distributed training → Promote to production via Flyte.

Related back office ops cases