Suraj Desai

About Suraj Desai

I build AI systems that actually work in production - not demos, not prototypes, not wrappers around an API.At Digg, I took content moderation from 50% recall to 95% using a RAG-based LLM pipeline I designed and shipped end-to-end. I built a hybrid search system combining BM25 and semantic embeddings that hits 80% MRR. I reduced duplicate content on a live platform by 80% using vector-based similarity detection. These are not side projects - they are systems running at scale.I have published research on knowledge distillation for dense retrieval and temporal fairness degradation in recommendation algorithms. I built a transformer from scratch off the GPT paper because I needed to understand the math, not just call the API.I work at the intersection of LLMs, retrieval systems, and agentic infrastructure. I ship and move fast, and own problems end-to-end.Publications: https://scholar.google.com/citations?view_op=view_citation & hl=en & https://github.com/suraj1297Articles: https://medium.com/@surajdesai1297/

Experience

1. Fouding Data Scientist

   Digg

   Dec 2024 — Present · Los Angeles, CA, US

   Developed a RAG-based AI agent for spam and toxic content detection, improving recall from 50% to 95% compared to third-party moderation tools, by implementing a retrieval-augmented pipeline that leveraged contextual document retrieval and LLM-based classification.• Built a hybrid search system with an MRR of 80% that improved relevance and retrieval accuracy for user queries, by combining BM25 lexical ranking with semantic embedding retrieval and storing vector representations in a VectorDB for efficient similarity search.• Reduced duplicate content on the platform by 80%, as measured by duplicate post frequency, by implementing a vector-based semantic similarity detection system that identifies near-duplicate content.• Developed a statistical social media post-ranking (trending) algorithm that increased the diversity of promoted content by 25%, by leveraging user interaction signals (likes, dislikes, comments) and applying fairness-aware weighting to ensure equitable visibility for underrepresented communities• Improved production reliability by implementing structured-logging and service-level metrics using Grafana Dashboards which reduced incident resolution time by 70%

Education

• Northeastern University

  Master of Science - MS, Information Systems

• University of Mumbai

  Bachelor's degree, Computer Engineering

  2015 — 2019