About

I’m a software and machine-learning engineer and researcher at Iowa State University, pursuing a B.S. in Data Science and an M.S. in Mechanical Engineering, and conducting research at the Translational AI Center (TrAC) with Prof. Soumik Sarkar.

My M.S. thesis builds agentic vision-language systems for cyber-agriculture; I co-first-authored SAGE, accepted at a CVPR 2026 workshop. I build end-to-end systems — from web apps and data pipelines to large-scale LLM/VLM evaluation and reproducible benchmarks — spanning agriculture, healthcare, and finance.

Three questions organise most of what I do:

  • Agriculture — does it still work in a real field? A model that scores well on a curated benchmark can fail on the next season’s photographs. I build agents that diagnose crop disease by reasoning over symptoms rather than pattern-matching pixels (SAGE, CVPR-W’26, on 335 crops and ~839K images), benchmark what vision-language models can actually do on plant stress (AgEval, WACV’25), ship retrieval-grounded pesticide-label guidance that abstains rather than cite the wrong herbicide (AgAdvisor, in beta with the Iowa Soybean Association), and measure what a year of distribution drift actually costs — 43.9 macro-F1 points (COMPAS’26).
  • Biology & health informatics — is the model reading the biology, or the batch? High accuracy on omics and pathology data is easy to get for the wrong reasons. I audit whether frozen pathology encoders carry real biological signal or just section identity (TabPFN probing), put three pathway-informed oncology architectures on identical folds and find most of their reported gaps vanish inside the confidence intervals (TRAPS, ICCA’26), and build multi-agent evaluation loops that cut ethical violations in medical LLM output by 89% (arXiv, ACM-BCB’25).
  • Optimization & geometry — what if the constraint is the interesting part? The best structure often falls out of what you are not allowed to do. When federated clients run different frozen encoders whose coordinate systems cannot be aligned, sending only coordinate-invariant Gram matrices recovers a global geometry in a single round at orders of magnitude less communication (FedQ-Vision). When compute is the binding constraint, routing each token through an adaptive number of recursions through one shared block beats bigger models with 75% fewer parameters (bioMoR) — and read as dynamical systems, transformers start to explain their own internals (APS 2026).

Representative work is highlighted with an orange badge. Click any title for the paper. Conference tiers shown in parentheses are CORE rankings; emerging venues not yet rated by CORE are marked top-tier by reputation.

Research

I work on agentic vision-language systems, large language models and multi-agent systems, and rigorous LLM/VLM evaluation, with applications across agriculture, healthcare, and finance. Representative work is highlighted. Click any title for the paper. Conference tiers shown in parentheses are CORE rankings; emerging venues not yet rated by CORE are marked top-tier by reputation.

Under Review & In Preparation

Under Review2026
Manuscript under review, 2026. (first author)
project page code

Federated learning across clients running different frozen vision encoders, whose coordinate systems cannot be averaged or aligned. Each client sends only the Gram matrix of its class prototypes — a coordinate-invariant summary — and the server recovers a global quotient geometry in a single round, with no alignment maps and no representation training. Across six benchmarks under heterogeneous CNN and transformer encoders it gave the best accuracy of all methods, +9.4 points over no collaboration, at one to two orders of magnitude lower communication.

Under Review2026
Manuscript under review, 2026. (second author)
project page

The first Mixture-of-Recursions framework for gene- and pathway-level omics learning: biological structure drives embedding smoothing, a structural attention bias, and a graph-aware router that sets each token’s recursion depth. Across eight single-cell and multi-omics benchmarks under one five-fold protocol it improved macro-F1 by 8.2 points over the strongest biology-agnostic MoR baseline while using 75% fewer parameters and up to 58% fewer FLOPs than a non-recursive Transformer.

In Prep2026
Manuscript in preparation, 2026. (fourth author)
project page

A training-free framework that audits whether frozen pathology foundation models encode biologically meaningful signal, pairing histology with spatial transcriptomics (HEST-1k, 240 samples across breast, skin, and brain) and using TabPFN as a standardized probe. Across five encoders, UNI led pathway decodability at 0.303, but organ-level shift cost 62% of the pathway signal and section identity stayed linearly decodable up to 214× chance — while benign augmentations retained ≥95% performance.

Conference, Journal & Workshop Papers

IROS2026 · Workshop
In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2026 (CORE A), Pittsburgh, PA — Workshop on From Innovation to Adoption: Hardware, AI, Policy, and Funding Challenges in Agricultural Robotics. (Accepted · poster, Oct 1, 2026 · invited to submit an extended version to the Springer Special Issue on Agricultural Robotics)

AgriMorph treats a field robot’s physical shape as a decision made while it works, not a fixed hardware property. Separate latent models of the terrain and of the robot’s own state let a learned world model roll out candidate futures under each available body, so shapes are compared before slip, stall or crop contact happens and changed only when the predicted gain beats the transformation cost. An ensemble-disagreement confidence gate blocks shape changes made on weak sensing, and one shape-conditioned controller carries over to every body without retraining.

WM@Booth2026
In World Models @ Booth (WM@Booth), University of Chicago Booth School of Business, 2026. (Accepted · poster, Aug 31–Sep 2, 2026)

A walk-forward backtesting framework benchmarks classical, neural, and world-model predictors on a global equity dataset, isolating when engineered features outperform raw price inputs across model families.

CVPR2026 · Workshop
In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 (CORE A*) — Workshop on Emerging Directions in Data for Multimodal Foundation Models. (Accepted · co-first author · M.S. thesis)
project page

An autonomous, symptom-grounded vision-language agent for crop-disease diagnosis, built on the largest plant-disease image–symptom dataset to date (335 crops, 1,251 classes, ~839K images). Stepwise diagnostic reasoning raised accuracy by +16.2 points and generalizes to unseen crops with no retraining.

AI Scientist Workshop2026 · Poster
In AI Scientist Summer Workshop, Microsoft Research New England, Cambridge, MA, 2026. (Selected for poster presentation · co-first author)

Selected for poster presentation at the AI Scientist Summer Workshop hosted by Microsoft Research (Aug 4, 2026): an autonomous, symptom-grounded vision-language agent for crop-disease diagnosis, evaluated on the largest plant-disease image–symptom dataset to date with stepwise diagnostic reasoning that generalizes to unseen crops.

ICCA2026
In 4th International Conference on Computing Advancements (ICCA), ACM Proceedings, 2026. (Accepted)
ArXiv project page

The first unified benchmark for pathway-guided cancer therapy-response modeling (BINN, GraphPath, PATH), over 2,622 TCGA patients across 5 cohorts; the best model reached 0.92 AUROC on prostate targeted-therapy prediction at 11% class prevalence.

IEEE COMPAS2026
In 3rd IEEE International Conference on Computing, Applications and Systems, IEEE Xplore, 2026. (Accepted · corresponding author)
project page code

A cross-year benchmark for rice-disease recognition on Bangladeshi field images: training on 2021–2025 and testing only on held-out 2026 data cost 43.91 macro-F1 points across all eight model/strategy configurations, with only 48.28% of validation performance retained. The better fine-tuning strategy proved architecture-dependent, and the most temporally robust model was not the one validation ranked first.

arXiv2026
arXiv preprint, 2026. Finalist, Paper Competition — 2026 IEEE Nexus Region 4 Annual Student Leadership Conference (SLC), Chicago, IL (Apr 2026)
ArXiv

A multi-agent iterative-refinement system (generators DeepSeek R1 + Med-PaLM; evaluators LLaMA 3.1 + Phi-4) aligned to AMA ethics and a five-tier safety assessment. Cut ethical violations by 89% with a 92% risk-downgrade rate over 900 clinical queries.

AAAI2026 · Oral
Intelligent Health Intervention System for Syndemic Management Among University Students: An AI-Driven Approach Using Reinforcement Learning
In Association for the Advancement of Artificial Intelligence (AAAI) 2026 Spring Symposium Series, San Francisco, CA. (Oral presentation)

A reinforcement-learning intervention engine that models intervention timing and selection as a sequential decision problem for syndemic health management.

APS2026
Interpreting Transformers through the Lens of Physics and Dynamical Systems
In American Physical Society (APS) Global Physics Summit, 2026. (first author)

A first-authored analysis interpreting transformer architectures through physics and dynamical-systems theory to explain their internal computation dynamics.

NE Agents / ACM-BCB2026
An Agentic Framework for Mechanistic Therapeutic Reasoning in Oncology via Pathway-Grounded Tree-of-Thought Inference
In North East AI Agents Day 2026 (Accepted); ACM Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM-BCB) 2026 (CORE B), Posters.

An agentic, pathway-grounded tree-of-thought inference pipeline that structures LLM reasoning over biological pathway knowledge for mechanistic therapeutic reasoning in oncology.

ICERIE2025
LlamaHealth: Assessing LLaMA’s Performance in Multi-Class Healthcare Test Result Prediction
In International Conference on Engineering Research, Innovation and Education (ICERIE), Shahjalal University of Science and Technology, Bangladesh, 2025. (Poster)

Benchmarked LLaMA on multi-class healthcare test-result prediction, assessing reliability for responsible medical-AI deployment.

AIChE2025
Benchmarking Bioemu-1: Establishing a Baseline for Protein Structural Ensemble Prediction
In American Institute of Chemical Engineers (AIChE) Annual Meeting, Boston, USA, 2025. (Poster)

Baseline benchmarks and reproducible metrics for protein structural-ensemble prediction in computational structural biology.

AIChE2025
Emissai: LLM-Powered Analysis of Carbon Footprint
In American Institute of Chemical Engineers (AIChE) Annual Meeting, 2025. (Presentation)

An LLM-powered tool to estimate and analyze carbon footprint from operational data.

MICCAI ODIN2025
In Oral and Dental Image Analysis (ODIN) Workshop @ Medical Image Computing and Computer Assisted Intervention (MICCAI) 2025 (CORE A). (Poster)
OpenReview

A privacy-preserving federated-learning framework for multi-modal 2D/3D dental imaging with cross-modal prompt alignment, hierarchical optimization, and Byzantine-resilient aggregation with differential privacy.

ACM-BCB2025
In ACM Conference on Bioinformatics, Computational Biology, and Health Informatics (ACM-BCB), Pennsylvania, USA, 2025 (CORE B). (Poster)
ACM DL

The poster prototype of the multi-agent medical-LLM safety framework (DeepSeek R1 / Med-PaLM with LLaMA 3.1 / Phi-4 evaluators), later extended into the arXiv study above.

WACV2025
In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025 (CORE A).
ArXiv

AgEval, a 12-task plant-stress phenotyping benchmark for zero-/few-shot in-context learning of VLMs (Claude, GPT, Gemini, LLaVA). Best-model F1 improved 46.24% → 73.37% (8-shot); strategic example selection added +15.38% F1.

JBIM2025
In Journal of Business & Industrial Marketing, 40(1), 265–280, 2025.
DOI project page

A GPT-4o pipeline over 430 Shark Tank pitches quantifying how market orientation and brand storytelling drive investor decisions, with expert annotation and statistical validation for a peer-reviewed B2B marketing study.

Projects

Deployed Systems

Built with real-world industry partners.

Live DemoHF Spaces
Live public demo on Hugging Face Spaces. (Translational AI Center / AIIRA · Iowa Soybean Association beta)
Demo
  • Built and deployed a multi-agent RAG assistant that answers pesticide-label questions grounded in CDMS labels, live as a public demo. Stack: FastAPI + LangGraph multi-agent orchestration, OpenAI embeddings + Qdrant vector search, GPT-4o, Streamlit/Gradio UI.
  • Industry connection: developed in the Translational AI Center (AIIRA) and beta-tested with the Iowa Soybean Association, iterating directly on grower and agronomist feedback (wrong-herbicide bias, answer verbosity, mobile confidence badge, cold-start latency).
  • Engineered faithful, abstaining retrieval — a product-disambiguating catalog plus abstention so the agent refuses rather than answer from the wrong herbicide label, with page-level citations: a concrete reliability guarantee for regulatory agrochemical guidance.
  • Productionized with account authentication and per-user persistent history, per-user daily quotas to protect shared API keys, response caching, and a concurrency-safe shared vector store; a 62-test suite runs in CI.

Experience

  • Graduate Research Assistant, Translational AI Center, Iowa State University — Jan 2025 – Present
  • Undergraduate Research Assistant, Translational AI Center, Iowa State University — Apr 2024 – Present
  • Undergraduate Software Developer, Dept. of Agricultural & Biosystems Engineering, Iowa State University — Jun – Dec 2024
  • Undergraduate Research Volunteer, Dept. of Computer Science, Iowa State University — Apr – May 2024

Advising & Mentoring

Honors & Awards

In the News

February 2026 Discovery and curiosity come together in undergraduate research — LAS News, Iowa State University. Featured for undergraduate research at the Translational AI Center on explainable AI for persuasive digital content, mentored by Soumik Sarkar and Priyanka Jayashankar.
October 2025 Award winners at the 22nd Annual Norman Borlaug Lectureship Poster Competition — Global Resource Systems, Iowa State University. Named among Iowa State students recognized at the competition, where I won first place in the undergraduate division.

Academic Service

Professional Memberships

Program Committee

Ethics Review

  • Conference on Neural Information Processing Systems 2026 (CORE A*) — Main Conference and Evaluations & Datasets Tracks
  • Conference on Neural Information Processing Systems 2025 (CORE A*) — Main Conference and Datasets & Benchmarks Tracks

Conference Reviewer

Workshop Reviewer

Journal Reviewer