Abhijit Chakraborty

I study how AI agents retrieve from data and tools they don't own, and how they decide when not to answer.

  • Manager, Partner Solutions Architect, Cloud Partnerships at MongoDB. I run the technical side of MongoDB's North America relationship with AWS and work with the Azure and Google Cloud teams.
  • PhD candidate at Arizona State University's CoRAL Lab, advised by Vivek Gupta.
Portrait of Abhijit Chakraborty
A question goes to one of three sources: tool catalogs, enterprise databases, or other organizations' data. The system checks what it retrieved, then answers with sources or abstains. A question from a person or an agent Tool catalogs SYNAPSE, Typed Fed. Artifacts Enterprise databases JoinGR, DeALOG Other organizations' data Federated RAG survey Check the output OSCAR, CaRE, EviSearch Answer with sources Abstain or escalate ongoing work A question goes to one of three sources, the system checks the result, and then answers with sources or abstains. A question from a person or an agent Tool catalogs SYNAPSE, Typed Fed. Artifacts Enterprise databases JoinGR, DeALOG Other organizations' data Federated RAG survey Check the output OSCAR, CaRE, EviSearch Answer with sources Abstain or escalate ongoing work
Most of my papers study one step of this loop. An agent chooses where to look among sources it doesn't control, checks what it retrieved, and then answers with sources or hands the question back. Paper names show where each one sits. Point at a source to trace its route, or select any box to read about that step.

News

At AWS re:Invent 2026 I'm co-presenting a breakout session on agent state and memory with Amazon Bedrock and MongoDB Atlas.
New preprint: JoinGR finds the tables a text-to-SQL question never names by walking the database's join graph.
Posted Typed Federated Artifacts for the Agentic Web, with Ni Trieu and Vivek Gupta.
EMNLP 2026 accepted two of our papers: DeALOG, and a survey of trajectory self-feedback in diffusion language models.
OSCAR will appear at COLM 2026.
We demoed JobMatchAI in the ACL 2026 System Demonstrations track.
TimeGate was presented at the CATS workshop at ICML 2026.
Our Federated RAG mapping study appeared in Findings of EMNLP 2025.

Research

Most enterprise AI systems depend on data and tools that other teams or organizations own. My papers ask what agents can share in that setting without moving raw data, how they find the right source, and how they check what comes back. I came to the PhD after fifteen years of building data platforms.

Retrieval and tool routing across distributed sources

Where should an agent look, and what can agents share?

Federated retrieval is where this started. Our mapping study in Findings of EMNLP 2025 covers the work from 2020 to 2025 that combines federated learning with retrieval-augmented generation.

SYNAPSE and Typed Federated Artifacts carry the idea over to tool routing. Agents share what they have learned about which tool suits which request, packaged as schema-checked objects instead of model weights. That lets agents running different frozen LLMs use the same knowledge. On StableToolBench the federated setup was within 1.1 points of a centralized baseline.

Even inside one database, the right table can be hard to find. JoinGR follows foreign-key joins to reach tables a question never mentions, and on the BEAVER enterprise benchmark it raised perfect recall@15 from 0.339 to 0.645. In DeALOG (EMNLP 2026), specialist agents for text, tables and images coordinate through a shared log, with no central controller.

Checking model output

How does a system check its own answer, and when should a person take over?

OSCAR (COLM 2026) looks for hallucinations inside diffusion language models. It runs several denoising chains in parallel, treats the positions where they disagree as suspect, and rewrites those positions using retrieved evidence.

Evaluation needs the same scrutiny. For CaRE we re-ran seven remasking strategies for masked diffusion models on twelve models under matched compute. Sampling temperature explained most of the variation (η² = 0.91). Once compute was matched, the strategy rankings reversed.

EviSearch keeps a person in the loop. It links every value it pulls from a clinical trial paper to its source and sends uncertain cells to a reviewer. On its own it was 91.7% accurate; after a reviewer checked 15.6% of the cells, accuracy reached 95.2%.

My current work is on selective prediction for knowledge graph completion: when a model can certify a prediction, and when it should abstain.

Structured knowledge

How much model does a knowledge graph need?

For link prediction, less than recent work assumes. RelatE is a real-valued phase-modulus embedding that matches or beats complex-valued and deep models on standard benchmarks (MRR 0.521 on YAGO3-10). It also trains 24% faster than RotatE and uses 22% less GPU memory.

JobMatchAI (ACL 2026 System Demonstrations) uses knowledge graphs and semantic search to match candidates to jobs, and explains each match.

Running ML in production

Why do models stall between prototype and production?

The MLOps mapping study (CSCE 2024) catalogs what stops ML projects in the data, model-building and deployment pipelines, and recommends a tool-agnostic fix for each. TimeGate (CATS workshop, ICML 2026) adds time-boxed promotion gates for models retrained continually on a fixed compute budget.

Publications

Peer-reviewed papers are marked. Google Scholar has the same list with citation counts.

2026

2025

2024

Track record

Outside the lab I have spent nearly two decades building teams and closing enterprise work, from data engineering at British Telecom to cloud partnerships at MongoDB.

Revenue and savings

  • $4.9Mthree-year deal

    US telecom customer that adopted Atlas as the embedding and memory layer for its Amazon Bedrock applications.MongoDB

  • $1.25Mclosed

    Media and entertainment project on Atlas.MongoDB

  • $1.2Min six months

    New account for Cisco's newly acquired WebEx, delivered by an India team I built.SRS Consulting

  • $750Ksaved per year

    Moved reporting from Business Objects to Oracle BI and CRM; reports also ran 15% faster.British Telecom

  • $250Ksaved per year

    Led the Oracle-to-Teradata migration for Cisco analytics.Infosys Consulting

Teams built and led

  • 2 → 20in twelve months

    Grew the delivery team (3 onshore, 17 offshore), which led to the onsite role running the account.British Telecom

  • 0 → 10in six months

    Hired the India team for Cisco WebEx. Nobody left.SRS Consulting

  • 20 + 4engineers, managers

    Led the group delivering OCI Autonomous Database work and a serverless point-of-sale system.Oracle, next-generation OCI

  • 3 regionsenablement

    Took train-the-trainer partner enablement from North America to India and ANZ; Vietnam is in discussion.MongoDB

Experience

2022 – now

Manager, Partner Solutions Architect, Cloud Partnerships MongoDB

I own the technical relationship with AWS across North America and work with the Azure and Google Cloud teams. The job is to design the architectures, run the enablement and shape the joint plays that turn a partnership into named accounts and deals we can trace.

  • An earlier proposal to put MongoDB in the short-term memory of Amazon Bedrock AgentCore was turned down. I reframed it around long-term memory, and AWS's AgentCore team then invited MongoDB in as one of the first ISV partners in the memory layer.
  • Co-presenting an accepted AWS re:Invent 2026 breakout on agent state and memory. It is the first breakout approved for members of MongoDB's partner team.
  • Built a claims-processing agent demo and workshop (S3 Tables and Aurora for retrieval, Bedrock for inference). MongoDB and AWS partner leadership approved it, and I presented it at MongoDB .local New York.
  • Account work includes a healthcare supply-chain customer that ranked among North America's top revenue models two quarters running, a global payment-gateway migration across North America, EMEA and APAC, and an app-modernization win against an AWS-first pitch.
  • Much of the role is leading people who don't report to me. I automated the AWS Skill Badge integration that AWS marketing now uses at events, run a biweekly technical partner community with members in India, LATAM, Singapore and APAC, recruited AWS technical champions, and wrote the field playbook for partnership deals.
  • Field leaders in the West, Canada, East and enterprise regions invite me to present partner wins in their team meetings. Atlas feature requests I raised with product management are now prioritized.
Ongoing

Co-founder JobMatchAI

An explainable job-matching venture built on CoRAL Lab research, with Mayank Vyas and Vivek Gupta.

2015 – 2022

Senior Manager and Engineering Lead, Next-Generation OCI Oracle

  • Joined as an individual contributor building DevOps tooling and distributed data platforms for Oracle Cloud Applications. Automated ETL across millions of rows, which cut manual work by 33%, and kept pipelines from five transactional sources at 99.9% uptime.
  • Moved to the next-generation OCI team and led 20 engineers and 4 managers on OCI Autonomous Database work and a serverless point-of-sale system.
  • Built the integration framework from Vertica on AWS to Salesforce Marketing Cloud (Lambda, EC2, S3, SQS) and the real-time Informatica MDM to Salesforce API behind the automated marketing campaigns. Designed the division's BI warehouse on Data Vault and Kimball models.
2014 – 2015

Senior Data Engineer Lead, Freight Data Platform ProKarma for Union Pacific

Took on a new line of business to open up freight operations data. Driver, linesman and track events streamed into Vertica at 2,000 events a second through Oracle GoldenGate change data capture. I also moved legacy ColdFusion reporting to Oracle Sales CRM and built the real-time monitoring front end.

2011 – 2014

Lead Business Value Consultant, then Engagement Lead SRS Consulting and Infosys Consulting, for Cisco and AC Nielsen

  • SRS Consulting, a startup serving Cisco, asked me to set up an India team for the newly acquired WebEx. Within six months the team had ten people, the account passed $1.2M in revenue, and nobody had left.
  • Infosys Consulting then hired me into its business value consulting practice, where I grew the manufacturing and analytics domains and brought in new clients, including AC Nielsen and Jaguar Land Rover.
  • Led the Oracle-to-Teradata migration for Cisco analytics, saving $250K a year in maintenance. Also cut vendor-data pipeline errors by 7%, sped up BI queries by 10%, and localized reports, which raised BI adoption by 7%.
2007 – 2011

Data Engineer, then Delivery Lead British Telecom

  • Grew the delivery team from 2 to 20 people in twelve months and moved onsite to lead and grow the account.
  • Led the move from Business Objects to Oracle BI and CRM, which saved $750K a year and made reporting 15% faster. Built the ETL platform for web apps with 10,000 daily active users ($85K a year saved in vendor costs) and the Orange Wholesale BI warehouse, which scaled to 115,000 daily users.

Talks

  • AWS re:Invent 2026 breakout session, accepted
  • MongoDB .local San Francisco and New York
  • AWS GenAI Loft
  • Mastercard CEO and CTO fireside chat part of an AWS campaign
  • LangChain community event
  • US Foods hackathon 38 hands-on builders from 102 registrants

Service

Conference reviewing

  • NeurIPS2025, 2026
  • ICLR2025, 2026
  • ICML
  • AAAI2025, 2026
  • ACL2026
  • EMNLP2025, 2026

Journal reviewing

  • Neurocomputing

Workshops

  • SurgeLLMProgram committee
  • DocuInsightProgram committee
  • Agentic Web, NeurIPSReviewer

Other

  • Harvard Business ReviewBoard of Advisors
  • Apache Software FoundationContributor

Education

  • PhD, Computer Science, Arizona State UniversityCoRAL Lab, advised by Vivek Gupta. Dean's List.
  • MS, Computer Science (Big Data Systems), Arizona State UniversityGold Medal, summa cum laude
  • BTech, Electrical Engineering, West Bengal University of Technology (MAKAUT)

Certifications

Oracle Certified Associate (Database Management); Oracle Certified Specialist (Database Development and Optimization); Oracle PartnerNetwork Certified Specialist (BI Technologies).

Contact

Research collaborations, talks, and partner work on AI with MongoDB Atlas.

Email abhijit.chakraborty@asu.edu for research and student projects. For partner or customer conversations, reach me on LinkedIn.