Mukundan M

At Swish, I productionized an ensemble ETA service using CatBoost, XGBoost, Optuna, H3 geospatial features, and Redis, achieving ≤2-minute MAE on 48% of orders and ≤4 minutes on about 80%.

I replaced Typesense with a fine-tuned MPNet and BERT-reranker search stack serving about 150 QPS, cutting null searches by 50%. I also built segment-aware ranking and cart upsell systems that improved add-to-cart conversion.

At CuriousInc (Tell M), I architected a canonical grocery catalog across seven platforms and made completeness measurable through sitemap-driven weekly diff-and-fill. I also redesigned agent execution into tool contracts and specialized subagents while improving ingestion failure visibility.

I build ML and platform systems end to end, from retrieval and recommender systems to PostgreSQL-backed taxonomy, load testing, and infrastructure. My work includes LegalRAG, a local six-agent contract-analysis pipeline with grounded citations and evidence-based refusal.