Jun 2022 – Present · 3 yrs 10 mos · Remote

• Traffic concentrated in SEO; Tasked to grow engagement and retention of TOI's owned platforms; AI we ship is seen by audiences without humans in the loop (requires high accuracy). Personalization doubled feed CTR & grew retention by 10%.
• Machine Learning:
• In-house data wasn't reliable; Built CDP, redid data pipeline, and drove adoption; We now process 100m events daily, and warehouse stores 750 GB. CDP is the backbone for ML.
• Website & apps designs added bias into data; Collaborated to redesign TOI Apps.
• Website & apps can't serve ML; Worked with engg. to drop cache & use servers to render.
• News is low ARPU business and ML is costly; Cost-modeled ML and optimized API to 195ms
• Stakeholders doubted if ML can run audience-facing tasks; Shipped transparent dashboards in Metabase & Kibana to track outcomes and ML Ops.
• Media struggle with retention ‘coz they run many revenue models at the same time; Proposed a method to mix ML with P&L to maximize revenue within target retention.
• Generative AI:
• Even if outcomes are high, does the feed pass the smell test? Increasingly building evals (labeled data) to track if the right stories are amplified. Using LLM to build new features.
• LLMs can summarize but in doing so they lose news peg; Building agents to catch news peg — why a story exists and using it to rewrite headlines to drive engagement.
• Piloted chatbot inside of TOI Apps to test their reaction to chat as an interface
• Conceptualized Context Cards — Amazon X Ray but for news
• Stack: LangChain, DSpy, GPT APIs, Perplexity APIs, Pine Cone, Embeddings

Dec 2021 – May 2022 · 5 mos · Remote

• TOI's editorial platform had been built for SEO and print legacy — not for Machine Learning. Before any of those could work, the foundations needed rebuilding. That was this role. I worked directly with the Editor-in-Chief across two interconnected problems.
• Taxonomy. TOI had no clean content classification system. Stories were filed by section (India, World, Sports) inherited from print, with no consistent logic underneath. Personalization and ML require clean taxonomy as features. We built one from scratch: 242 nodes across 4 levels, discovered through grounded theory with editorial leads, designed to be both human-taggable today and AI-classifiable tomorrow. The result: audiences could finally see the diversity of what TOI covered. Editorial got clear signals on which topics performed. The ML team got clean training data.
• First recommendation algorithm. Combined the new taxonomy with Signals, our story-ranking system, to build TOI's first recommendation widgets (Read Next, Up Next, Utility Belt). Result: 107% higher CTR than the previous widget. Subscribers clicking through at 8-9% on TOI+ content.

Dec 2020 – Nov 2021 · 11 mos · Remote

• TOI's revenue was concentrated in display programmatic ads. I was brought in as editorial product lead to relaunch TOI+. My focus was discovering what people would actually pay for, and building the product around that answer.
• I worked directly with the Business Head and Editor-in-Chief, bringing strategy, editorial judgment, and product into the same room.
• TOI had a structural tension at its core: a newsroom built for wide, accessible city coverage and a mandate to launch an exclusive subscription product. "Exclusive" and "accessible" pull in opposite directions. Resolving that tension was the first and most important product decision.
• Beyond content strategy, we invented new formats to make the product worth subscribing to: Newscards (snackable visual stories) and CardHD Immersives (rich, long-form multimedia narratives). Additionally, we doubled down on data driven storytelling.
• Outcome: 120,000 annual paying subscribers in year one. 90%+ of published stories triggered at least one subscription.