Dinkar Juyal

I am an ML Engineer at Meta working on agentic systems. Previously, I was a Staff ML Researcher and Engineer at PathAI, focused on foundation models and computer vision for digital pathology.

My work spans foundation models for pathology (PLUTO), domain generalization (ContriMix, best-performing official submission on the Camelyon17 WILDS leaderboard), interpretability (Additive MIL, sparse-autoencoder analysis of pathology embeddings), and evaluation of AI agents as ML engineers (ML-Dev-Bench). This work has been applied clinically, including AI-based automation of enrollment criteria and endpoint assessment in liver disease trials (Nature Medicine, 2024). See publications for the full list.

I write about ML research — mostly notes on agents, foundation models, and evaluation. New posts go up on Substack first (see below); some are later mirrored here.

Talks & articles

ongoing projects

A few things I'm actively building and experimenting with:

CDBench

Controlled multi-fault debugging benchmark: injects N simultaneous bugs into real open-source library code and measures fix rate as N scales from 1 to 10. A GRPO-trained 7B model with iterative repair closes part of the gap to a 32B dense ceiling.

tiny-vlm-lab

A minimal, NanoChat-style vision-language lab: synthetic grid-world scenes with exact ground truth, fast CPU training, and explicit checks for whether a small VLM is actually grounding language in the image rather than guessing from text alone.

benchmark-creator

Turn any GitHub repo into a behavioral benchmark for LLMs in one command: reads the library's README, extracts behavioral rules, and runs an adversarial two-player game to find execution-verified questions where a model knows a rule but misapplies it.

writing on substack

Recent essays are published on Substack before anywhere else — that's where to find the newest posts.

Curiosity, Inspiration and Exploration thumbnail

Curiosity, Inspiration and Exploration

Adding curiosity as a trait in agentic systems.

August 29, 2026

Verification Loops, Documented Context thumbnail

Verification Loops, Documented Context

The two things that matter for E2E agentic AI for knowledge work.

February 16, 2026

Evaluation for Agentic AI thumbnail

Evaluation for Agentic AI

How benchmarks and evaluation strategies have changed as we move from LLMs to agents, and what lies ahead.

January 25, 2026

latest posts

Aug 29, 2026 Curiosity, Inspiration and Exploration
Feb 16, 2026 Verification Loops, Documented Context
Jun 01, 2025 How to think with images

selected publications

  1. PLUTO: Pathology-Universal Transformer
    Dinkar Juyal, Harshith Padigela, Chintan Shah, and 30 more authors
    arXiv preprint arXiv:2405.07905, 2024
    Accepted at ICML 2024 workshops: ML for Life and Material Science, Foundation Models in the Wild, and Accessible and Efficient Foundation Models for Biological Discovery
    arXiv
  2. ContriMix: Scalable Stain Color Augmentation for Domain Generalization without Domain Labels in Digital Pathology
    Tan H. Nguyen, Dinkar Juyal, Jin Li, and 10 more authors
    arXiv preprint arXiv:2306.04527, 2023
    Best-performing official submission on the Camelyon17 track of the Stanford WILDS leaderboard
    arXiv Code Website
  3. ML-Dev-Bench: Comparative Analysis of AI Agents on ML Development Workflows
    Harshith Padigela, Chintan Shah, and Dinkar Juyal
    In ICLR 2025 Workshop on Deep Learning for Code (DL4C), 2025
    arXiv Code
  4. AI-based Automation of Enrollment Criteria and Endpoint Assessment in Clinical Trials in Liver Diseases
    Janani S. Iyer, Dinkar Juyal, Quang Le, and 28 more authors
    Nature Medicine, 2024
    HTML