Hi, I’m Vignav.

I'm an undergrad studying computer science at Harvard. Broadly, my academic interests center around natural language processing and multimodal learning in embodied agents.

I currently hold research scientist positions at MIT CSAIL, Harvard's Kempner Institute for AI, and Harvard Medical School, where I am co-advised by Profs. Pulkit Agrawal, Martin Wattenberg, and Pranav Rajpurkar. Previously, I built Arc, a tooling layer for enterprises to provision teams of specialized LLM agents (deferred $1M in funding to continue undergrad).

I have 5 years of software experience, 10+ hackathon wins, and have published several papers in top-tier AI/ML venues. In my free time, I enjoy playing tennis and practicing the double bass.

Press: Geekwire (x2), Mercury News, CA Times, East Bay Times, CBS News, Business.org, AP News, Central Charts, Markets Insider, Morningstar, GlobeNewswire, KFM BFM, Euroinvestor, OneNewsPage, MarketsAsk, CompuServe

Experience

Improbable AI Lab @ MIT CSAIL

Research Scientist

2024 -

Kempner Institute for AI @ Harvard SEAS

Research Scientist

2024 -

  • Training LM agents via policy learning to play coordination games by communicating via activations; achieves SOTA performance with far less compute

Medical AI Lab @ HMS DBMI

Research Scientist

2022 -

  • Developed LM-based NER models (>1M Huggingface downloads) to remove references to priors in chest X-ray radiology reports (work cited by Google, DeepMind, & Microsoft Research)
  • Leveraging unsupervised domain adaptation and co-training methods to develop foundation models for zero-shot medical image segmentation

QIAI Lab @ Stanford

Research Intern

2020 - 2022

  • Built novel algorithm to compute chest X-rays as coronal projections of axial CT volumes; trained Mask R-CNN to segment COVID lung lesions on real patient CXRs using computed CXR dataset (+64.5% IOU over SOTA)
  • Developed FundusNet, a contrastive learning framework using neural style transfer for detection and classification of referrable vs. non-referrable diabetic retinopathy with up to 90% label reduction; outperformed SOTA AUC by 9.6%

Arc

Founder & CEO

2023 -

Deferred $1M in funding, currently on hiatus

  • Built custom infrastructure for agent auto-specialization, synthetic dataset generation, and multi-agent orchestration
  • Piloting tech w/ Alchemy (Series C1 web3 development platform); partnered with Superpowered AI (YC S23) to develop SOTA retrieval-augmented generation systems

Speakeasy

Cofounder & Co-CEO

2022 -

  • AI startup providing automated pronunciation feedback and accent training services for enterprise
  • Built first AI model that measures mouth/tongue position purely from user audio; signed LOIs w/ top hospitality brands (Four Seasons, Marriott, Hyatt, Fairmont, IHG, etc.)

Publications

* first-author

Improving Radiology Report Generation Systems by Removing Hallucinated References to Non-existent Priors*

ML4H (NeurIPS) ’ 22, SAIL ’ 23

Contrastive learning-based pre-training improves representation and transferability of diabetic retinopathy classification models

Nature Scientific Reports

Unsupervised Tokenization Learning

EMNLP ’ 22 (Oral)

COVID-19 Lung Lesion Segmentation Using a Sparsely Supervised Mask R-CNN on Chest X-rays Automatically Computed from Volumetric CTs*

SIIM ’ 21 (Oral)

Context-Driven Question Answering Based on Link Grammar*

AGI ’ 21 (Oral)

Interpretable Natural Language Segmentation Using Link Grammar*

SAIence ’ 20 (Spotlight)

+4 more

Patents

[US #17/447,508] SYSTEMS AND METHODS FOR EVALUATING GAME ELEMENTS

2024

[US #63/156,357] UNSUPERVISED NATURAL LANGUAGE GENERATION USING LINK GRAMMAR FOR GENERAL CONVERSATIONAL INTELLIGENCE

2021

[US #63/156,359] INTERPRETABLE NATURAL LANGUAGE SEGMENTATION BASED ON LINK GRAMMAR

2021

Projects

+ more on GitHub

See LinkedIn, GitHub, and my resume for additional details.

© Vignav Ramesh, 2024.