Home
Résumé (PDF, opens in a new tab)
← All work
Professional workAI platforms & infrastructure

Better creative workflows. Shared decisions. Infrastructure that keeps up.

At Rakuten, I lead product work across shared decisioning, multi-agent creative workflows, and the infrastructure that supports them.

View résumé (PDF) ↗

Product flow

  1. 01Shared campaign + marketplace signals
  2. 02Recommendation, ranking + pacing
  3. 03Common APIs + integration contracts

The takeaway

Connecting AI orchestration, shared platform contracts, and operational performance.

Recommendation response time
−20%
Creative revision cycles
−30%

Sanitized reconstruction from role materials. The metrics below come from separate Staff PM initiatives and are not presented as one experiment.

The problem

Recommendation, ranking, and budget-pacing logic had grown across separate systems. Creative generation and review were also fragmented. As platform scale increased, product teams needed shared decisions, clear contracts, and infrastructure that could keep up.

My ownership

As Staff Product Manager, I led product architecture across three connected initiatives: a common decisioning layer, a multi-agent creative workflow, and the cloud and data infrastructure supporting them.

I defined shared signals, evaluation metrics, APIs, integration contracts, orchestration, memory and retrieval behavior, model evaluation, access and retention controls, and technical requirements for containerized services.

A decisive product moment

Should we reduce online state, or scale the live decisioning architecture?

Precomputing more decisions offline would reduce the number of values stored in Redis and simplify the online path. It would also narrow what the live decisioning layer could respond to as the platform grew.

Move more decisions offline

Precompute more of the answer and store less live state in Redis, reducing pressure on the existing architecture.

Shard the Redis architecture

Chosen

Distribute decisioning state horizontally so the platform could preserve its online requirements as scale increased.

Why

I drove the sharded architecture because the product still needed low-latency online decisions. The system constraint had to change without removing the live product capability it supported.

What changed

The migration improved p95 latency from 4 seconds to 300 milliseconds and reduced infrastructure costs by 30%.

System view

How the Staff PM scope fits together

A portfolio-level view of three separate initiatives, rather than a single runtime request path.

  1. 01Shared decisioningCommon campaign and marketplace signals → recommendation, ranking, and pacing → shared APIs and evaluation metrics
  2. 02GenAI creative workflowAgent orchestration → memory and retrieval → model evaluation and review → access and retention controls
  3. 03Infrastructure foundationContainerized services across AWS and GCP → Docker and Kubernetes → sharded Redis architecture
Foundation

Outcome lens · recommendation response time · product decision velocity · creative revision cycles · p95 latency · infrastructure cost

Leadership scope

How I made the work move

My role was to create shared decisions across the people building, evaluating, operating, and using the product.

01Engineering
Converted product needs into platform abstractions, API and integration contracts, infrastructure requirements, and migration outcomes.
02Data science
Defined shared signals and evaluation measures so ranking, recommendation, and pacing could be assessed consistently.
03Creative workflow stakeholders
Translated a fragmented generation-and-review journey into orchestration, memory, retrieval, evaluation, and governance requirements.
04Product teams
Used common interfaces to reduce repeated decision work while preserving the needs of each consuming product.

Selected product artifacts

The structure behind the decisions

Sanitized reconstructions show how I framed scope, contracts, and measurement without exposing internal systems or customer data.

Shared decisioning contract

Define what each consuming product could rely on.

Inputs
Common campaign and marketplace signals
Decisions
Recommendation, ranking, and budget pacing
Interface
APIs, integration contracts, and evaluation metrics

GenAI workflow specification

Treat AI output as a governed product workflow.

Orchestration
Agent roles, sequence, and handoffs
Context
Memory, retrieval, and data access
Quality & control
Model evaluation, review behavior, and retention rules

Professional outcomes

What changed

  • p95 infrastructure latency improved from 4s to 300ms; infrastructure costs fell 30%.
  • Multi-agent GenAI orchestration reduced creative revision cycles by 30%.
  • Shared decisioning reduced recommendation response time by 20% and increased product decision velocity by 25%.
Technical context
  • LangChain
  • Memory & retrieval
  • Model evaluation
  • Docker & Kubernetes
  • AWS & GCP
  • Sharded Redis

Up next

WhatsApp Catch-Up ↗

Let’s talk about the product decisions behind the work.

Get in touch ↗