Amazon Seller Assistant
Defined the science roadmap and designed the core multi-agent architecture behind Amazon's seller-facing assistant, launched at Accelerate 2024.
Applied Research Scientist / Amazon
Harsha Aduri
I lead science and architecture across Seller Experience and Seller Agentic Intelligence.
My work spans multi-agent systems, LLM tool use, model post-training, RAG, and multimodal experiences. Outside work, I build research tools and an occasionally over-engineered home lab.
01 / Selected impact
Research translated into working systems
Selected examples of applied research carried through architecture, production, and measurable outcomes.
Text2API
Built an agent that selects and executes more than 50 APIs with 96% accuracy, before native LLM tool calling became standard.
Canvas
Designed the code-generation architecture behind a multimodal Seller Assistant experience that turns conversations into dynamic, personalized visual workspaces for exploring business data and taking action.
Production impact
Delivered more than $200M in operational savings across ML automation initiatives and large-scale decision systems.
02 / Selected work
Research that reaches the real world
I work across research, engineering, and product constraints, with an emphasis on systems that remain useful after the demo.
Paper Radar
An AI-curated feed for keeping up with ML research. It discovers candidate papers, ranks them against my research interests, classifies topics, and publishes concise analyses.
- Seed- and query-based discovery across active ML research areas
- Heuristic filtering plus Claude-assisted relevance ranking
- Topic classification and abstract-grounded analysis
- Static publishing for a fast, searchable reading queue
CURATION PIPELINE
ACTIVE
- 01 Discover Semantic Scholar seeds and targeted queries
- 02 Rank Heuristics and Claude relevance scoring
- 03 Classify Research topics and content type
- 04 Analyze Abstract-grounded synthesis and signal score
- 05 Publish Static pages with search and filters
03 / Research & IP
Current research and intellectual property
Selected work on deep research, RAG evaluation, and generative interfaces. Status is shown explicitly.
DR+Tools: Tool-augmented Deep Research in Real World Applications
Rafael Ferreira, Harsha Aduri, Flavio Di Palo
RAGU-Bench: An End-to-End User-Quality Benchmark for Retrieval-Augmented Generation Systems
Harsha Aduri, Bhanu Teja Rangaraju
Beyond Factual Grounding: The Case For Opinion-Aware Retrieval-Augmented Generation
Aditya Agarwal, Harsha Aduri, Alwar Nakkiran
LLM-Compiled Executable Pipelines for Refreshing and Lazy Loading Generative User Interfaces
Pending patent application
04 / Current threads
What has my attention
Applied research across language, reasoning, agents, and human-centered seller experiences.
Tracking how better context, tools, and evaluation change what models can reliably do.
3D printing, PC thermals, and home automation. A suspicious amount of this is about airflow.