NACHIKET KATHOKE

Ads Recommendation and Ranking Unified Ranking & Personalization Systems

• Led as the primary developer of personalized recommendation and ranking models (GBTs, multi-head NNs), optimizing CTR, CVR, and partner ROI under strict latency constraints.
• Designed a multi-objective scoring framework using linear scalarization across AOV, distance, and user-intent signals, with weights tuned via Bayesian optimization, delivering ₹1 improvement in unit economics (UE).
• Launched dynamic ad serving based on click distribution analysis, stabilizing ad performance and recovering production regressions, yielding approximately 10 bps CVR uplift (approximately 6K iOPD).
• Implemented personalized dynamic ads ranking, improving relevance, monetization trade-offs, and achieving ₹2 RPO uplift without degrading organic metrics.

Unified Ranking & Personalization Systems

• Led a platform initiative to build unified Ads and Organic recommendation models, enabling legacy system consolidation while maintaining key business metrics.
• Designed and executed experiments on 1.1B+ sessions and 41M+ users, optimizing training strategy, labeling, loss functions, and Learning-to-Rank architectures using ordinal funnel labels (impression → click → order), and robust negative sampling.
• Developed and iteratively improved GBTs, multi-head NNs, sequential personalization models (Transformer-based), and FT-Transformer architectures to capture short-term intent and handle high-cardinality categorical features.
• Achieved approximately 10% NDCG uplift in ranking quality while preserving AOV and last-mile delivery performance, enabling the safe unification of multiple recommendation pipelines.

Multimodal Agentic Video Generation & Content Discovery

• Designed an end-to-end GenAI content generation pipeline, combining vision models and LLM-based prompt enrichment, converting static menu images into animated 9:16 high-conversion discovery assets.
• Implemented trust-aware ranking with probabilistic de-boosting to safely scale AI-generated content, balancing novelty and user conversion and discoverability.
• Built near real-time ingestion pipelines using Kafka and RILL, enabling freshness and high-throughput content delivery into discovery surfaces.