Hybrid ML + heuristics infrastructure serving contextually relevant help to millions of users in real time.
Company Google
Year 2021–2022
Role Tech Lead, Personalization Platform
MillionsUsers served with personalized help daily
Real-timeSignal integration across product surfaces
Privacy-firstCross-product signal architecture
The Challenge
Replacing Generic Help with Context-Aware Personalization
Vikram built a real-time personalization platform that surfaces contextually relevant help to millions of users daily — eliminating unnecessary support contacts at scale.
Help surfaces across the product ecosystem were serving generic content — the same articles to every user regardless of context, history, or current issue. This created unnecessary support contacts, agent overhead, and user frustration when the answer was already available but never surfaced at the right moment.
Building a personalization platform for help content presented distinct challenges: strict latency constraints on a real-time serving system, complex multi-round privacy approvals for cross-product signal access, and multi-system signal integration complexity across disparate product surfaces. High-traffic scaling requirements demanded rigorous capacity planning and load testing before every launch.
"Personalization at this scale isn't a model problem — it's an infrastructure problem. The model is the easy part."
System Architecture
Four-Component Personalization Platform
Component 1
Personalization Service
Java-based backend defining APIs, models, and core components. Real-time signal integration for live context-aware recommendations.
Component 2
ML Recommendation Infrastructure
Feature extraction pipelines feeding TensorFlow training workflows. Servo-based model serving for low-latency inference. Interaction logging integration for continuous model improvement.
Component 3
Signal Integration Layer
Cross-product signal API federation with authentication adaptations. Scalability optimization for high-traffic signal consumption.
Component 4
Personalized Content Service (PCS)
Rule config system enabling configurable targeting logic without model retraining. Third-party API integration with SLO-compliant infrastructure.
Key Innovations
Hybrid ML + rule-based architecture combining model intelligence with configurable business logic
Cross-product signal federation enabling help surfaces to draw context from across the full product ecosystem
Privacy-compliant architecture meeting regulatory and policy requirements without sacrificing targeting granularity
Outcomes & Impact
Better Help, Fewer Contacts, Lasting Foundation
MillionsUsers served with relevant help daily
The platform improved user satisfaction by surfacing more relevant help content at the right moment in the user journey. Reduced support costs followed directly from fewer unnecessary contacts reaching live agents. The long-term personalization foundation established by this project enabled scalable content targeting across channels, creating compounding value as the signal library and model quality improved over time.
My Role
Privacy, Architecture, and Engineering Leadership
Drove the privacy design process, navigating regulatory and policy requirements to unblock cross-product signals
Co-authored key architecture documents defining the personalization platform's long-term design
Mentored junior engineers through build, testing, and production launch phases
Explainer Video
Project Overview
Tech Stack
Tools & Technologies
ML / AI
TensorFlowServo (ML Serving)
Backend
JavaRule EnginesConfig Systems
Platform
Cross-product Signal APIsInternal Backend Frameworks