Hardening Agents for E-commerce Scale: From RL Alignment to Reliability // Panel 2

36 snips

Dec 2, 2025

Guest

In this engaging discussion, expert panelists share insights into the world of e-commerce agents. Arushi Jain, a Microsoft applied scientist, delves into post-training techniques that enhance AI reliability for tasks. Swati Bhatia from Google Cloud talks about using Direct Preference Optimization to fine-tune support routing. Audi Liu from Inworld AI discusses architectural trade-offs in voice models for better accuracy. Isabella Piratininga of iFood highlights personalization challenges in Brazil. Together, they explore the complexities of automating customer interactions and the future of AI.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

ANECDOTE

DPO+PEFT Fixed A Broken Support Router

Swati Bhatia described an agent that misrouted complex support tickets until a DPO+PEFT adapter fixed compliance rapidly.
The team raised policy adherence from 45% to 90% within 48 hours using human-reviewed training data.

INSIGHT

Post-Training Replaced Prompt Engineering

Arushi Jain said post-training has replaced prompt engineering for many Copilot features at Microsoft.
She credits grounded post-training on Microsoft data for reducing variance and hallucinations in production.

ANECDOTE

Cascading Beat End-To-End Speech Models

Audi Liu shared a production improvement when switching from end-to-end speech models to a cascaded STT+LLM+TTS architecture.
The cascaded setup increased tool-calling accuracy and customer retention despite ~200ms extra latency.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

Thanks to Prosus Group for collaborating on the Agents in Production Virtual Conference 2025.

Abstract //

The discussion centers on highly technical yet practical themes, such as the use of advanced post-training techniques like Direct Preference Optimization (DPO) and Parameter-Efficient Fine-Tuning (PEFT) to ensure LLMs maintain stability while specializing for e-commerce domains. We compare the implementation challenges of Computer-Using Agents in automating legacy enterprise systems versus the stability issues faced by conversational agents when inputs become unpredictable in production. We will analyze the role of cloud infrastructure in supporting the continuous, iterative training loops required by Reinforcement Learning-based agents for e-commerce!

Bio //

Paul van der Boor (Panel Host) //

Paul van der Boor is a Senior Director of Data Science at Prosus and a member of its internal AI group.

Arushi Jain (Panelist) //

Arushi is a Senior Applied Scientist at Microsoft, working on LLM post-training for Computer-Using Agent (CUA) through Reinforcement Learning. She previously completed Microsoft’s competitive 2-year AI Rotational Program (MAIDAP), building and shipping AI-powered features across four product teams.

She holds a Master’s in Machine Learning from the University of Michigan, Ann Arbor, and a Dual Degree in Economics from IIT Kanpur. At Michigan, she led the NLG efforts for the Alexa Prize Team, securing a $250K research grant to develop a personalized, active-listening socialbot. Her research spans collaborations with Rutgers School of Information, Virginia Tech’s Economics Department, and UCLA’s Center for Digital Behavior.

Beyond her technical work, Arushi is a passionate advocate for gender equity in AI. She leads the Women in Data Science (WiDS) Cambridge community, scaling participation in her ML workshops from 25 women in 2020 to 100+ in 2025—empowering women and non-binary technologists through education and mentorship.

Swati Bhatia //

Passionate about building and investing in cutting-edge technology to drive positive impact.

Currently shaping the future of AI/ML at Google Cloud.

10+ years of global experience across the U.S, EMEA, and India in product, strategy & venture capital (Google, Uber, BCG, Morpheus Ventures).

Audi Liu //

I’m passionate about making AI more useful and safe.

Why? Because AI will be ubiquitous in every workflow, powering our lives just like how electricity revolutionized our society - It’s pivotal we get it right.

At Inworld AI, we believe all future software will be powered by voice. As a Sr Product Manager at Inworld, I'm focused on building a real-time voice API that empowers developers to create engaging, human-like experiences. Inworld offers state-of-the-art voice AI at a radically accessible price - No. 1 on Hugging Face and Artificial Analysis, instant voice cloning, rich multilingual support, real-time streaming, and emotion plus non-verbal control, all for just $5 per million characters.

Isabella Piratininga //

Experienced Product Leader with over 10 years in the tech industry, shaping impactful solutions across micro-mobility, e-commerce, and leading organizations in the new economy, such as OLX, iFood, and now Nubank. I began my journey as a Product Owner during the early days of modern product management, contributing to pivotal moments like scaling startups, mergers of major tech companies, and driving innovation in digital banking.

My passion lies in solving complex challenges through user-centered product strategies. I believe in creating products that serve as a bridge between user needs and business goals, fostering value and driving growth. At Nubank, I focus on redefining financial experiences and empowering users with accessible and innovative solutions.