Researcher & Builder

Falak
Pabari

I build models and products that work when biology is noisy, data is messy, and assumptions break.

Brown CS + Applied Math, Honors. Heading to Columbia in Fall 2026 to work on probabilistic models for single-cell genomics.

Research → CV GitHub
latent space  ·  VAE embedding
live
↓
About

I grew up in Jamnagar, a small town in Gujarat where getting a proper diagnosis meant driving hours to a city. That stayed with me. I'm interested in systems that go to people, not the other way around, and in models that are upfront about what they can't do.

At Brown, I work under Professor Ritambhara Singh and Professor Ying Ma on two questions: can structure emerge from data that was never labeled for it? And: when a model is wrong, does it know?

Outside the lab, I play chess badly and dance less badly, and love to talk about books and food. I've lived in three countries and counting!

In Fall 2026, I'm heading to Columbia to work on hierarchical generative models for single-cell data.

Falak Pabari

Research

Singh Lab, Brown University Honors Thesis  ·  April 2026

Deep Learning Models for Robust Latent Structure in Single-Cell and Methylation Data

Pabari F., Advised by Singh R.

Epigenetic clocks commit to linearity: fit a line between methylation and age, keep the high-slope sites, discard the rest. Aging is not a smooth slope. Using a hybrid VAE and Contrastive encoder on the Hannum cohort (656 samples, 470k CpGs), two GMM clusters in the learned latent space capture 53.8% of SNITCH-classified nonlinear CpGs at odds ratios of 4.97 and 2.74. The CTCF/NF1 transcription factor axis identified in Module 9 independently replicates in Grolaux et al. (2026) on EPICv2 data from a different cohort and an entirely different methodological framework. The central claim: nonlinear epigenetic aging structure is encoded in trajectory shape geometry, not in the statistical relationship between methylation and age.

Read Thesis →
Singh Lab, Brown University Poster  ·  Expected 2026

Contrastive Disentanglement of Methylation Variation Reveals Population-Structured Aging Trajectories

Pabari F., Singh R.

Contrastive learning reveals a hierarchy in methylation variation (ancestry, sex, then age), clarifying why standard epigenetic age prediction requires strong supervision and why cross-cohort clocks degrade. Presented at scverse Conference (Stanford, 2025) and the Broad ML for Drug Discovery Symposium.

View Poster →
Ma Lab, Brown University Manuscript in Preparation  ·  Expected 2026

Uncertainty-Aware Detection of Perturbation Escapees in Large-Scale CRISPR Screens

Pabari F., Lee J., Ma Y., Chang S.

In a Perturb-seq screen, most cells respond to a genetic perturbation as expected. Some do not. The question is not just which cells escape, but whether your model knows they are escaping or confidently mislabels them. A dual-encoder VAE with z/s disentanglement, adversarial regularization, HSIC independence penalty, and Wasserstein-based escape modeling separates perturbation-specific signal from background biological variation. Across 67 perturbations: 98.6% classification accuracy, 11 escape candidates identified, CEBPE surfaced among top salient gene loadings. Poster and oral presentation at UTRA Research Symposium (2025).

View Poster →
Humans to Robots Lab, Brown University CogSci 2024  ·  RSS 2024

Find It Like a Dog: Using Gesture to Improve Robot Object Search

He I., Pelgrim M., Lee K., Pabari F., Buchsbaum D., Tellex S., Nguyen T.

Gesture-guided object search on Boston Dynamics Spot. Evaluation across 72 trials on 6 human-dog pairs using Euclidean distance, weighted accuracy, and perplexity. 83% accuracy. Proceedings of the Cognitive Science Society (CogSci 2024); presented at Robotics: Science and Systems (RSS 2024), Delft, Netherlands.

Read Paper →

Projects

Co-founder 2025 – Present

Skiwi

Skin conditions change over time. Skiwi tracks them. Built on the premise that longitudinal pattern detection beats one-shot confidence, especially on consumer photo data where lighting, angle, and skin tone vary every session. React / Node / Python + FastAPI, multi-model inference pipeline with YOLOv8 and GPT-4V.

2,000+ waitlist 300+ users
Visit skiwi.ai →
Computer Vision 2026

Vera

Burn and wound classification across four public datasets. The evaluation framing that actually matters: per-Fitzpatrick-skin-type bias analysis. Fine-tuned EfficientNet and ConvNeXt classifiers, U-Net segmentation, Grad-CAM saliency, longitudinal wound-area tracking.

View Repo →
Full-Stack 2026

Atlas

Real-time geopolitical event graph. Fetches live global news via GDELT, clusters it geographically, renders an interactive world map with causal edge analysis between events.

Visit Live →
Genomics / ML 2024

GeneExpress

Fine-tuned Nucleotide Transformer (2.5B parameters, multi-species) to predict tissue-specific gene expression across 218 tissues and estimate variant-driven expression changes. A probe of how much conserved regulatory structure survives cross-species transfer even when numeric accuracy is modest.

View Repo →
Full-Stack 2025 – Present

Connect2

Invite-only networking marketplace connecting students with professionals at top companies. Admins curate and approve profiles; students browse, filter, and book sessions directly. Next.js 15 / TypeScript / Supabase (PostgreSQL + RLS) / Vercel, with magic link auth and automated booking request management.

View Live →

Writing

Under maintenance

Posts coming soon.

Let's Collaborate

If any of this resonates (the research, the questions, or just the chess losses), I'd love to hear from you.

Get in Touch View Resume