The Data Exchange with Ben Lorica Why Digital Work is the Perfect Training Ground for AI Agents
10 snips
Sep 18, 2025 
Andrew Rabinovich, CTO and Head of AI at Upwork, dives into the synergy between digital work marketplaces and AI agents. He reveals how Upwork's unique Reinforcement Learning from Experience (RLEF) methods optimize agent training. The discussion highlights the importance of user evaluation in shaping AI performance and explores innovative multimodal feedback systems. Rabinovich also envisions a future where human workers and AI coalesce seamlessly, providing expedited services and new opportunities for freelancers.
AI Snips
Chapters
Transcript
Episode notes
UMA Orchestrates Digital Work End-to-End
- Upwork built UMA, a meta-agent that facilitates client–freelancer interactions end-to-end on the platform.
- UMA helps clients express needs, finds matching freelancers, and orchestrates contracts and work.
Mixture Of Fine-Tuned Skills With RAG
- Upwork layers fine-tuning: a base personality for UMA and many narrow skill models for specific tasks.
- They use RAG to bring dynamic platform data into UMA's skills that need up-to-date context.
Route For Latency; Use Narrow Models
- Prioritize latency-sensitive routing and reserve heavy reasoning only for tasks that need it.
- Train narrow, task-specific models instead of relying on general reasoning models for real-time interactions.