Machine learning & agentic AI

Viresh Jivane.

Principal ML Engineer,
Agentic AI

PayPal · Venmo

I lead agentic AI experiences for Venmo. Previously, I worked on Instagram ranking optimization at Meta and led machine learning teams at Walmart.

Recognitions

INFORMS

Franz Edelman Laureate

Named a Franz Edelman Laureate for deep reinforcement learning and multiobjective markdown optimization across Walmart’s U.S. stores.

The project was selected among six global finalists. INFORMS describes the award as the “Nobel Prize” of operations research and advanced analytics.

View INFORMS recognition

Experience

PayPalVenmo

Principal ML Engineer, Agentic AI

2026–present

Leading Venmo’s agentic AI experiences.

Agent architecture
Planning & routing
Agent & tool orchestration
Context engineering
Agent evaluation
Guardrails
Prompt optimization
Agent observability
Latency optimization

MetaInstagram

ML Engineer

2024–2026

Ranking optimization for Instagram.

Binary stochastic neurons
Gradient balancing
Frustration modeling
Pacing optimization

WalmartApplied AI

Senior Manager II / Principal ML Engineer

2019–2024

Built and led ML engineering teams.

Deep reinforcement learning
Demand forecasting
Language models
Fulfillment prediction

Earlier

Intuit

Large-scale Apache Spark streaming and batch pipelines for tax filings.

2015–2019

HSBC

Financial reconciliation and quality engineering.

2009–2014

MS, Software EngineeringSan José State University

BE, Information TechnologyPune Institute of Computer Technology (P.I.C.T)

Publications & patent

Google Scholar

INFORMS Journal
on Applied Analytics

2021

A Multiobjective Optimization for Clearance in Walmart Brick-and-Mortar Stores (opens in a new tab)

Coauthored research combining deep reinforcement learning, simulation, and optimization for markdown pricing across Walmart’s U.S. stores.

Read the paper 18 citations

Scientific Data, Nature Portfolio · 2022

The United States COVID-19 Forecast Hub dataset (opens in a new tab)

Contributed forecasts to the U.S. COVID-19 Forecast Hub as a consortium member.

Read the paper 229 citations

U.S. patent application · July 2026

Systems and Methods for Predicting Fulfillment Times and Updating User Requests (opens in a new tab)

Co-inventor · US 2026/0212285 A1

View the application

Articles

Bayesian ensembling and uncertainty

Quantifying Prediction Uncertainty with Bayesian Model Fusion (opens in a new tab)

Model 150% weight

Cat 90%Dog 10%

Model 250% weight

Cat 10%Dog 90%

Weighted sum0.50 × 0.9 + 0.50 × 0.1 = 0.50

CombinedPrediction

Cat 50%Dog 50%

1.00bits of predictive entropy

Each model is confident, but they disagree. The mixture is uncertain.

Bayesian model averaging. Posterior weights come from model priors and marginal likelihoods. Entropy measures uncertainty: 0 bits is certain; 1 bit is evenly split here. Toy probabilities.

Bandit-based hyperparameter tuning

Efficient Hyperparameter Tuning with Successive Halving (opens in a new tab)

Validation accuracy Retained× Stopped

Eight configurations receive increasing cumulative training budgets. Only retained configurations continue.
Config.	1epoch	2epochs	4epochs	8epochs
A	78 percent, retained	—	—	—
B	63 percent, stopped	—	—	—
C	80 percent, retained	—	—	—
D	67 percent, stopped	—	—	—
E	74 percent, stopped	—	—	—
F	81 percent, retained	—	—	—
G	72 percent, stopped	—	—	—
H	84 percent, retained	—	—	—

8 × 1configs × epochs each

8 configurations evaluated. Keep 4; increase each survivor’s budget to 2 epochs.

Budgets are cumulative per configuration. Rank only on validation data. An early low score can discard a slow starter. Scores are illustrative.

Generalization to unseen classes

Zero-Shot Learning: Understanding Machines That Learn Like Humans (opens in a new tab)

New image

Predicted attributes

Equine shape 0.92·Striped coat 0.95

Compare with class descriptions

A simplified attribute matcher compares the image with candidate class descriptions.
Attribute	HorseSeen	TigerSeen	ZebraUnseen
Equine shape	0.94	0.18	0.93
Striped coat	0.08	0.91	0.96
Distance	0.87	0.74	0.01

ZebraClosest attribute match

The zebra description is closest to the image’s attributes, despite no labeled zebra images in training.

Learn an image-to-attribute mapping on known classes; supply descriptions for new classes at prediction time. This toy matcher uses two attributes and Euclidean distance. Real models use richer representations.

Visual explanations