Ashmit Mukherjee

Computer Science · NYU Abu Dhabi

Ashmit Mukherjee

I’m interested in computational social science, especially how AI advice affects decisions people make together.

I’m a Computer Science senior at NYU Abu Dhabi. I coordinate software development and review for cooperation experiments, and train and evaluate protein language models.

[email protected]GitHub ↗
Ashmit Mukherjee

Research

Two ongoing collaborations

With Ben Rosche & Hanan Salam · 2026–present

Human–AI collaboration

How does AI advice affect cooperation when individual and group incentives differ?

I work on the software behind these experiments: game rules, participant interfaces, AI advice, and the records researchers use for analysis. I coordinate development and review the implementation against the study design.

The software and my contribution

A cooperation experiment

  1. A shared dilemmaA choice that benefits one person can leave the group worse off.
  2. Advice and decisionsParticipants receive guidance and make their own choices.
  3. Individual and group outcomesThe study examines how advice changes cooperation over repeated interactions.
Simplified research setting. The project studies different forms of AI assistance; no human-study results are reported here.

Co-first author · eBRAIN Lab · 2026–present

Bio-LoRA & protein-model evaluation

Can biological information help adapt a protein language model?

I added signal-peptide prediction to the evaluation and ran comparisons across three ESM-2 model sizes. Follow-up feature-injection experiments produced a mixed result: an advantage at 650M did not carry across model sizes or establish a general benefit from biological features.

My contribution and comparison results

Follow-up experiment · ESM-2 650M

Six-method ESM-2 650M comparison. Mean exact cleavage-site match ranges from 83.64% for vanilla LoRA to 86.08% for BLOSUM62 feature injection. Error bars show sample standard deviation across three seeds.
All six methods in the saved 650M comparison. Means ± sample SD across three seeds; 1,231 signal-peptide-positive examples. Full fine-tuning led the separate 150M and 3B comparisons. Feature injection is separate from the manuscript’s initialization method.
Manuscript

Bio-LoRA: Biological Priors as Inductive Bias for Low-Rank Adaptation.
Co-first author. Unpublished manuscript, 2026.
About the manuscript and my contribution

Current direction

Synthetic Pilot Lab

My capstone examines how faults in experimental software distort recorded outcomes. The implementation pairs correct and faulty game engines, replays scripted decisions, and checks the resulting records with an independent evaluator.

Offline diagnostics and measurement analysis completed using synthetic data. Model-based fault discovery remains untested.

Other projects

FieldLens, a local tool for reviewing fictional incident records; Hinglish named-entity recognition; and CAMP, a team-built campus equipment application.

Projects and coursework

A little background

I’m majoring in Computer Science with a minor in Economics and expect to graduate in May 2027. I’m preparing for doctoral research and am interested in applying computation to social, economic, and historical questions.

I studied away at NYU Paris and NYU New York. On campus, I’ve facilitated workshops for first-year students and worked front of house at the Arts Center.

Previously, I was an ML engineering intern at Zeek (summer 2025), working on customer segmentation and model comparison.