Subhra Rath

Role Led data, AI, and decision-intelligence strategy for the APJ-wide program
Scale 7,000+ customers · 50+ partners · $36M+ ARR influence · >100% YoY growth

Problem. Manual eligibility checks, credit issuance, and account audits could not scale to an APJ-wide adoption program. Regional teams also needed consistent visibility into partner performance, executive reporting, and audit trails without slowing execution.

What changed. Spearheaded agentic AI automation across eligibility assessment, credit issuance, and account audits — with ML-led lead scoring, GenAI-powered account planning, and pre-eligibility assessment models for partner-led motions. Built the underlying operating model, unified reporting, and reusable execution patterns for scalable field delivery.

Outcome. Cut eligibility-check and credit-issuance processing time by >85%, scaled the program to 7,000+ customers and 50+ partners, and influenced $36M+ ARR with >100% YoY growth — while tightening governance, consistency, and audit readiness.

• AWS
• Agentic AI
• Automation
• GenAI
• ML
• GTM Analytics
• Executive Reporting

Role Led design and delivery
Scale 40+ countries · daily-refresh signal

Problem. Static pricing tables were creating margin leakage in some markets and demand suppression in others, with no consistent way to compare Airalo prices against local competitors or partner rate cards across regions.

What changed. Built a country-level pricing analytics workflow on AWS that combined partner rate cards, competitor price signals, and market-level elasticity views into a single decision layer — with experiment guardrails so pricing changes could be trialled in one market before wider rollout.

Outcome. Gave the commercial team continuous visibility into where each market sat versus competitors, supported more disciplined acquisition and margin decisions, and shortened the loop between a market-level pricing question and a defensible answer.

• Pricing
• Analytics
• AWS
• Experimentation

Role Designed and delivered churn propensity modelling
Scale Large-scale consumer loyalty base · monthly scoring

Problem. The CRM program treated customers uniformly and missed early defection signals — high-value loyalty members were lapsing before the team had a chance to intervene, and campaign investment was skewed toward customers already engaged.

What changed. Built a churn propensity model using engagement, redemption, purchase-recency, and promo-response features, scored monthly and pushed back into the CRM to drive segmented retention journeys instead of blanket sends.

Outcome. Identified high-risk members several weeks earlier than the previous rules-based flags, powered a high-ROI churn-intervention program, and delivered $1.2M incremental revenue at 12× ROI on the retention effort.

• Churn
• CRM
• Loyalty
• Predictive Modelling

Role Designed NLP / GenAI classification approach
Scale 15+ languages · app reviews, tickets, social

Problem. Customer feedback across markets could not be reliably interpreted using English-first sentiment tooling — non-English tickets, reviews, and social mentions were either mis-scored or dropped from the CX signal entirely, leaving regional issues invisible.

What changed. Designed a multilingual classification layer combining transformer-based models with an LLM step for theme extraction, so sentiment, complaint clusters, and emerging market-level issues could be tagged consistently across all 15+ supported languages without routing everything through a translation hop.

Outcome. Improved support triage on non-English queues, surfaced regional CX patterns that had previously been buried in aggregated dashboards, and gave marketing a shared language for reading customer sentiment across markets.

• GenAI
• NLP
• Multilingual
• Customer Intelligence

Role Designed and implemented support automation pattern
Scale High-volume repetitive support queries

Problem. Support tickets were growing faster than L1 capacity, while product, policy, and troubleshooting knowledge was scattered across help centre, internal wiki, and legacy ticket history — so agents spent as much time searching as answering.

What changed. Built a RAG-based assistant pattern using LangChain and a vector store over the consolidated knowledge base, with an evaluation harness for answer quality, source-citation guardrails to prevent hallucinated policies, and human-in-the-loop escalation for anything the retriever was not confident about.

Outcome. Reduced manual handling pressure in the target repetitive-query categories, gave agents faster access to source-cited answers inside their workflow, and surfaced concrete knowledge-base gaps driving escalations so the content team could close them.

• LLM
• RAG
• LangChain
• Support Automation

Role Led analytics / recommendation approach for campaign targeting
Scale Multi-market retail campaigns

Problem. Cross-sell and up-sell campaigns relied heavily on static rules and merchandiser intuition, so email and in-store campaigns kept over-serving the same customers with the same categories while missing higher-propensity next-basket combinations.

What changed. Built a hybrid recommendation approach — collaborative signals from purchase history combined with content-based product affinities — wired into campaign management, email journeys, and in-store customer touchpoints, with A/B holdouts so uplift could be measured campaign by campaign.

Outcome. Supported measurable campaign improvements across all three markets, including roughly 20% lift in target store footfall on featured campaigns, meaningful basket-size gains, and a ~20% cross-sell uplift on targeted segments.

• Recommendations
• Retail
• Growth
• Campaign Analytics

Role Built / led hybrid NLP pipeline
Scale High-volume onboarding and document routing

Problem. Manual document classification and PII extraction created turnaround delays, inconsistent routing, and operational bottlenecks in a regulated workflow where errors carry compliance risk, not just SLA cost.

What changed. Built a hybrid pipeline pairing a fine-tuned transformer classifier with a Named Entity Recognition model for structured field extraction, backed by a deterministic rule layer for edge cases and a review queue that routed low-confidence documents to human operators with a clear audit trail.

Outcome. Cut routine manual processing on the covered document types, improved routing consistency across intake channels, and gave the operations team defensible visibility into exception handling and reviewer throughput.

• NLP
• NER
• Document AI
• Operations