Aayush Agarwal
Senior LLM Research Engineer, Bloomberg

I'm a research engineer working on large language models and reasoning systems—currently at Bloomberg, previously at Google and Microsoft. Outside of work, I write essays at Intuition First to break down complex math and physics concepts into intuitive, learnable building blocks, and build software experiments on the side.

Engineering & Code · 4 projects

Software & Systems

Web apps, mobile tools, and software experiments built on the side.

  • Fiducia React Native
    A mobile application built with Expo and TypeScript for tracking and comparing investment portfolios.
  • PactPal TypeScript
    An accountability-driven task planner designed to curb procrastination through binding daily pacts.
  • Flash Cards Web App
    An interactive web application featuring curated flashcard decks for staff engineer interview preparation.
  • Culinary Canvas JavaScript
    A modern recipe gallery application with modular single-source definitions and modal views.
Highlights
— Shipped a retrieval-augmented earnings-call analyzer used by 40+ analysts
— Cut structured-extraction error rate by 30% via a fine-tuned adapter stack
Highlights

GenAI Reliability & Grounding: Solved the LLM hallucination problem for Search AI by building a real-time cross-verification engine with Knowledge Graph structured data; reduced contradictory answers by 90% across 100+ countries and 20+ languages while deploying large-scale caching to manage high-concurrency latency and compute costs.

Search Ranking & Ecosystem Expansion: High-quality creator content was getting buried by corporate SEO, so I built custom ML classifiers that cut brand noise 10x and scaled index/retrieval support across multi-platform feeds (X, Facebook, LinkedIn) by 3x, surfacing 2M+ authentic creators across 100M+ queries.

AI & Intelligent Workflows: Replaced legacy heuristic link recommendations in Google Docs by architecting a production Two-Tower neural network (increasing CTR by 40%+) and engineered the end-to-end distributed ML training/inference pipeline for Smart Compose, expanding it into 4 new global languages.

Highlights

Real-Time Event Question Answering: Engineered neural NLP extraction models to answer high-volume Bing queries on live, unfolding events by mining and synthesizing breaking news articles in real time, surfacing direct answers on the search results page.

Trending Entity Detection: Architected near real-time streaming pipelines over Bing query streams and news feeds to detect trending and breakout entities within minutes, powering dynamic discovery across Bing.com.

Knowledge Graph Semantic Grounding: Linked dynamic event and entity extractions into the Bing Knowledge Graph, establishing semantic entity resolution and structured answer cards for timely, high-velocity topics.