Portrait of Bhavdeep Singh Sachdeva

Software engineer · NLP researcher

Bhavdeep Singh Sachdeva

I turn scattered data into one view people trust.

  • 12+region launches
  • 10xaccounts per organization
  • 300+citations
  • 8publications
Explore the workGet in touch

Story

Security data at cloud scale. Research on what makes NLP benchmarks robust. Now building for the home.

AWSGuardDuty & Security Hub ResearchBenchmarks, data quality, robustness TesseraA verified home passport

AWS GuardDuty & Security Hub

Software Development Engineer, Amazon Web Services, Seattle · From Aug 2021

Many AWS accountsAWS Config snapshotsOrganize & processParquet filesDelta Lake tableMY TEAM · INGESTION & PROCESSINGDownstreamBuilt & indexedfor later searchCorrelate across accountsHow resources connectAccount AAccount BAccount CRelated items, one findingPublicbucketPermissiveroleExposedinstance1 findingPatterns identified by security engineers AWS accountsAWS ConfigsnapshotsOrganize & processParquet filesDelta Lake tableMY TEAM · INGESTION & PROCESSINGDownstreamBuilt & indexed for later searchHow resources connectAccount AAccount BAccount CRelated items, one findingPublicbucketPermissiveroleExposedinstance1 findingPatterns identified by security engineers
Illustrative concept with generic examples, not the production design

Finding risk in how resources connect

Led the ingestion and processing of AWS Config data, from S3 Parquet files into a Delta Lake, so it could be correlated across accounts.

Preview · 22 regions · Jun 2025GA · Dec 2025
Details

AWS Config records how each resource in an account is set up. I led the small team that built the pipeline that ingested that data, processed it, saved it to S3 as Parquet files and loaded it into a Delta Lake table, so configuration from many accounts could be analyzed together. That made correlation possible: seeing how resources and services in different accounts interact, and spotting combinations that are risky together but harmless alone, such as a public bucket, a permissive role and an exposed instance in related accounts. Security engineers identified many of these patterns and grouped related items into single findings. The data was then built and indexed downstream for later search. Part of the new AWS Security Hub: preview in 22 regions (Jun 2025), GA Dec 2025.

Illustrative positions
New-region deployment: months → hours
CloudFormation → AWS CDK

Months to hours

12+ region launches, 2022–2025.

Details

Cut GuardDuty new-region deployment from months to hours by migrating infrastructure from CloudFormation to AWS CDK and removing region-specific code paths, sequenced by service dependencies; delivered 12+ region launches (2022–2025).

Member accounts per organization: 5,000 → 50,000, enabled with one API call

10x scale

Built GuardDuty's AWS Organizations enablement service.

Launched Mar 2023
Details

Built GuardDuty's AWS Organizations enablement service, powering the Mar 2023 launch that enforces threat detection on all existing and new accounts with one API call, and scaling support 10x, from 5,000 to 50,000 member accounts per organization in 2023.

Research

Benchmarks, data quality and robustness in NLP.

BenchmarksData qualityNumerical reasoningRobustnessApplied ML
300+ TOTAL CITATIONS NumGLUE 200+

Google Scholar, as of Oct 22, 2026

  1. Real-Time Visual Feedback to Guide Benchmark Creation: A Human-and-Metric-in-the-Loop Workflow

    Anjana Arunkumar, Swaroop Mishra, Bhavdeep Sachdeva, Chitta Baral, Chris Bryan

    Proposed VAIDA, real-time visual feedback for crowdworkers that cut artifacts in created samples by 45.8%.

    Benchmarks · Data quality

    Read paper ↗

  2. NumGLUE: A Suite of Fundamental yet Challenging Mathematical Reasoning Tasks

    Swaroop Mishra, Arindam Mitra, Neeraj Varshney, Bhavdeep Singh Sachdeva, Peter Clark, Chitta Baral, Ashwin Kalyan

    8-task arithmetic reasoning benchmark; state-of-the-art language models score 46.4% below humans. 200+ citations (Google Scholar); used in training and evaluating math LLMs such as MAmmoTH.

    Benchmarks · Numerical reasoning

    Read paper ↗

  3. Generalized but not Robust? Comparing the Effects of Data Modification Methods on Out-of-Domain Generalization and Adversarial Robustness

    Tejas Gokhale, Swaroop Mishra, Man Luo, Bhavdeep Singh Sachdeva, Chitta Baral

    More data helps both OOD accuracy and adversarial robustness; data filtering can hurt OOD accuracy.

    Robustness · Data quality

    Read paper ↗

  4. DQI: A Guide to Benchmark Evaluation

    Swaroop Mishra, Anjana Arunkumar, Bhavdeep Sachdeva, Chris Bryan, Chitta Baral

    Proposed DQI, a data quality metric for benchmarks; showed spurious biases that AFLite filtering misses.

    Data quality · Benchmarks

    Read paper ↗

  5. Do We Need to Create Big Datasets to Learn a Task?

    Swaroop Mishra*, Bhavdeep Sachdeva*

    *Equal contribution

    RoBERTa nears full performance with 2% of SNLI; proposed a baseline to find a task-sufficient dataset.

    Data quality

    Read paper ↗

  6. Towards Question Format Independent Numerical Reasoning: A Set of Prerequisite Tasks

    Swaroop Mishra, Arindam Mitra, Neeraj Varshney, Bhavdeep Sachdeva, Chitta Baral

    Introduced NUMBERGAME, a benchmark evaluating numerical reasoning across eight diverse question formats.

    Numerical reasoning · Benchmarks

    Read paper ↗

  7. DQI: Measuring Data Quality in NLP

    Swaroop Mishra, Anjana Arunkumar, Bhavdeep Sachdeva, Chris Bryan, Chitta Baral

    Introduced DQI to help create bias-free datasets; models trained on DQI-renovated SNLI generalize better OOD.

    Data quality

    Read paper ↗

  8. Traffic state detection using smartphone based acoustic sensing

    Arshvir Kaur, Nitakshi Sood, Naveen Aggarwal, Dinesh Vij, Bhavdeep Sachdeva

    Journal of Intelligent & Fuzzy Systems 32(4): traffic state detection from acoustic data collected on smartphones.

    Applied ML

    Read paper ↗

Full list on Google Scholar ↗

Now building

Passport Owner 1Owner 2Owner 3

Tessera Build Inc., Co-founder (pre-launch)

Tessera Build Inc. is an early-stage company I co-founded. Its vision is an intelligence and orchestration layer for the home: a verified home passport that begins with the builder's records and stays with the house through every owner. The goal is for homeowners to set up services, see what's under warranty, and reach the right trade from one place, and for builders to spend less time on calls and email. The idea draws on my earlier work at AWS unifying security findings across AWS accounts into one trusted view.

Journey

Systems Engineer

Stealth, Unisys · Bangalore

Security logs + ML, two hackathon wins.

More
  • Built a proof of concept that ran malware samples in Cuckoo Sandbox and classified behavior from system and security logs with an ML model; won Fan Favorite at Unisys Winter Hackfest India (Dec 2017) and a hackathon Grand Prize (2018).
  • Delivered secure IPP support in Enterprise Output Manager 14.0.

Research Assistant

Cognition and Intelligence Lab, Arizona State University · Tempe

NumGLUE, DQI, VAIDA and robustness research.

More
  • Co-authored NumGLUE (ACL 2022), an 8-task arithmetic reasoning benchmark on which state-of-the-art language models trailed humans by 46.4%; 200+ citations (Google Scholar).
  • Co-authored a Findings of ACL 2022 study showing more data improves both OOD accuracy and adversarial robustness, while data filtering can hurt OOD accuracy.
  • Co-developed DQI (Data Quality Index) and co-designed VAIDA (EACL 2023), which reduced benchmark artifacts by 45.8%.

Software Development Engineer

GuardDuty / Security Hub, Amazon Web Services · Seattle

Security Hub ingestion, 12+ region launches, 10x org scale.

More

See the AWS section above.

Volunteer Research Contributor (unpaid)

Aya Open-Science Initiative, Cohere For AI · Remote

Punjabi-language data for Aya.

More

Curated and validated Punjabi-language data for Aya, Cohere For AI's open-science multilingual project, released as the Aya Dataset and Aya Model (arXiv, 2024).

Honors

  1. 2009

    Silver Medal · Government of India

    National Child Award for Exceptional Achievement

    Awarded at Vigyan Bhavan, New Delhi, with a Titan watch sent as a gift from Dr. Manmohan Singh, who wasn't there in person. Still one of my happiest memories.

  2. 2010

    Bronze · Team India roller hockey

    14th Asian Roller Sports Championship

    3rd place in Kaohsiung, plus a man-of-the-match nomination. It gave me the confidence that I can take on big problems.

  3. 2017–18

    Fan Favorite · Grand Prize

    Fan Favorite Award, Unisys Winter Hackfest India

    Dec 2017 · followed by a hackathon Grand Prize in 2018

    Ran malware in Cuckoo Sandbox and trained an ML model to spot vulnerability patterns: a 0-to-1 version of an idea I later saw at AWS scale in GuardDuty.

Education
  • M.S. (Thesis), Computer Science, Arizona State University, 2019–2021
  • B.E., Computer Science, Panjab University, Chandigarh, 2012–2016

Writing

AI Safety & Governance

Human-in-the-Loop Governance for Evolving AI Agents

2024 Multi-Agent Systems

Balancing Human and Model Capabilities

2024 AI Research

From Monolithic Models to Compound AI

2024 Multilingual NLP

Preserving Panjabi Through Aya

2024

Let's talk

Happy to discuss NLP research, AI safety and research collaborations.

bhavdeepsachdeva@gmail.com