Applied science leadership · Amazon AGI

Ankit Chadha

Director of Applied Science at Amazon AGI.
I lead science and engineering teams building grounded reasoning for AI agents, agent memory, and web-scale retrieval across text, images, and structured data.

• Google Scholar
• ORCID
• GitHub
• LinkedIn
• ACL Anthology
• OpenAlex
• Amazon Science
• X

I lead Applied Science and Engineering teams at Amazon AGI, turning research in foundation models and agents into systems that work at web scale. My work spans post-training and alignment, multimodal and grounded retrieval, agent memory, and evaluation. A recurring question connects it: how can an AI system find reliable evidence, reason over it, and retain useful experience? Previously I led deep learning work at Salesforce and worked at Apple and Samsung.

Selected research contributions

For the live publication and citation record, see my Google Scholar profile.

Agent memoryACL 2026 (Main Conference, Long Papers), pages 16470–16489+3.5 pp on LOCOMO

APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AI

Pratyay Banerjee, Masud Moshtaghi, Shivashankar Subramanian, Amita Misra, Ankit Chadha

APEX-MEM uses an append-only property graph of temporally grounded, entity-centric events and a multi-tool retrieval agent to resolve evolving information at query time.

Research contributionIntroduced an append-only property graph memory that represents conversations as temporally grounded, entity-centric events and resolves evolving information at retrieval time.

Agent memoryEMNLP 2026, accepted, to appear+31.7 pp on KGQAGen-10k

APEX-EM: Non-Parametric Online Learning for Autonomous Agents via Structured Procedural-Episodic Experience Replay

Pratyay Banerjee, Masud Moshtaghi, Ankit Chadha

APEX-EM introduced procedural knowledge graphs for LLM agent memory: its typed Procedural Knowledge Graph stores complete solution traces and retrieves them by meaning, operation structure, and graph traversal, with no model retraining.

Research contributionIntroduced procedural knowledge graphs for LLM agent memory through a typed Procedural Knowledge Graph, procedural-episodic experience replay, and cross-domain structural retrieval.

Federated learningNAACL 2022 (Main Conference), pages 2576–2586+1.83 BLEU on German to English

Training Mixed-Domain Translation Models via Federated Learning

Peyman Passban, Tanya Roosta, Rahul Gupta, Ankit Chadha, Clement Chung

Five domain-specific translation models use federated learning to build a mixed-domain neural machine translation system without pooling their data, with a method for selecting impactful parameters to control communication bandwidth.

Research contributionIntroduced a federated approach to mixed-domain neural machine translation and a bandwidth-control method that selects impactful parameters during federated updates.

Web-scale QACIKM 2022, pages 4707–4711+9.6 pp on WDRASS

WDRASS: A Web-scale Dataset for Document Retrieval and Answer Sentence Selection

Zeyu Zhang, Thuy Vu, Sunil Gandhi, Ankit Chadha, Alessandro Moschitti

WDRASS introduced a web-scale open-domain question answering dataset with 64,000 questions and 800,000+ labeled passages and sentences drawn from 30 million documents.

Research contributionIntroduced a web-scale dataset that connects document retrieval, passage reranking, and answer sentence selection using complete sentence-level answers rather than short answer matching.

All 13 papers, with abstracts and numbers →

Research guides

Practical reading guides that connect related papers without replacing their primary evidence.

• Agent memory: changing facts and reusable proceduresA guide to two complementary agent-memory designs: temporal event memory and procedural experience replay.
• Adaptive handoffs for multi-agent systemsA practical reading guide to choosing a typed dependency graph or natural-language handoff for each agent task.
• Web QA: retrieve evidence, rank answers, extend to more languagesA research lineage from web-scale training data to context-aware answer ranking and low-resource multilingual supervision.

Research and systems

Agent memory2026
2 papersCurrent

Introduced typed procedural knowledge graphs for reusable agent experience and append-only property graph memory for temporal conversational reasoning.

• APEX-MEM: Agentic Semi-Structured Memory with Temporal Reasoning for Long-Term Conversational AIACL 2026 (Main Conference, Long Papers)+3.5 pp on LOCOMO
• APEX-EM: Non-Parametric Online Learning for Autonomous Agents via Structured Procedural-Episodic Experience ReplayEMNLP 2026, accepted, to appear+31.7 pp on KGQAGen-10k

How agents delegate2026
1 paper

Introduced adaptive routing between typed dependency graphs and natural-language handoffs for multi-agent LLM systems.

• Routed Graph Handoff: Adaptive Format Selection for Multi-Agent LLM DelegationEMNLP 2026, accepted, to appear+12.7 pp on τ-bench retail

Web-scale question answering2022–2023
4 papers

Built web-scale document retrieval, passage reranking, and answer sentence selection methods spanning thirty million documents and nine languages.

• WDRASS: A Web-scale Dataset for Document Retrieval and Answer Sentence SelectionCIKM 2022+9.6 pp on WDRASS
• Cross-Lingual Knowledge Distillation for Answer Sentence Selection in Low-Resource LanguagesFindings of ACL 2023+12.1 pp on Xtr-WikiQA, Arabic
• Question-Context Alignment and Answer-Context Dependencies for Effective Answer Sentence SelectionINTERSPEECH 2023+6.07 pp on WikiQA
• Efficient Fine-Tuning Large Language Models for Knowledge-Aware Response PlanningECML PKDD 2023 (Research Track), LNAI 14170+5.1 pp on WikiQA

Learning from data that cannot move2021–2022
2 papers

Developed mixed-domain neural machine translation and communication-efficient federated learning without pooling data across domains.

• Training Mixed-Domain Translation Models via Federated LearningNAACL 2022 (Main Conference)+1.83 BLEU on German to English
• Communication-Efficient Federated Learning for Neural Machine TranslationNeurIPS 2021 ENLSP Workshop+1.06 BLEU on Federated translation

Steering what a model says2022–2023
2 papers

Developed methods for controlling toxicity, sentiment, simplicity, and viewpoint through hidden-state transformations and attribute conditioning.

• Controlled Text Generation with Hidden Representation TransformationsFindings of ACL 2023-0.74 toxicity on RealToxicityPrompts
• ACM -- Attribute Conditioning for Abstractive Multi Document SummarizationarXiv preprint+7.13 on MultiNews

Experience

1. 2021 – now

   Director of Applied ScienceAmazon, Amazon AGI

   Lead a team of scientists and engineers on grounded, multimodal language model systems: knowledge graphs, web and multimodal search, retrieval and agentic retrieval, and the post-training and evaluation behind them.

2. 2020 – 2024

   Advisor, AI programStanford Center for Professional Development

   Advised on Stanford's professional AI courses XCS221, XCS224N and XCS224U.

3. 2019 – 2021

   Senior Manager, Deep Learning Research and EngineeringSalesforce

   Built and led the team behind Einstein Intent, Einstein NER and Einstein Sentiment, the language understanding under Einstein Bots.

4. 2016 – 2019

   Previous companiesSamsung R&D America, Apple

Education

• MS, Computer EngineeringUniversity of Minnesota, Twin Cities
• Graduate AI coursework, CS221, CS224n, CS231n, CS234Stanford University