cv

MSc. (Hons.) in Biological Sciences & BEng. (Hons.) in Electronics & Instrumentation

BITS Pilani, K.K. Birla Goa Campus

Minor in Data Science in Climate & Health

• Linear Algebra
• Probability & Statistics
• Applied Statistical Methods
• Foundations of Data Science
• Machine Learning
• Object Oriented Programming
• Bioinformatics
• Operating Systems

AI Research Intern

CSIR — Central Electronics Engineering Research Institute (CEERI)

Brain-computer interface research using EEG signals and deep learning.

• Designed a lightweight attention-CNN for real-time EEG finger-movement detection — 85% accuracy, F1 0.75, inference <10 ms per segment (deployment-compatible)
• Shipped an interactive multi-channel EEG visualization dashboard for researchers exploring temporal correlations across electrodes

Undergraduate ML Researcher

APPCAIR, BITS Pilani

Transformer + LLM work for computational biology and drug discovery with Prof. Ashwin Srinivasan and Prof. Raviprasad Aduri.

• Designed LSRPI — Transformer that predicts residue-level RNA-protein interactions from primary sequence alone, no structural input required; validated via Gnina docking simulations (co-first-author paper under review)
• Built a semi-automatic molecule-generation + retrosynthesis pipeline with logical feedback loops on top of GPT-4 / Claude; shipped 20+ novel JAK2 and DBH inhibitor candidates to wet-lab queue
• Diagnosed and removed preprocessing bottlenecks in the prior XGBoost RPI model — 5% accuracy lift and 10% inference speedup

ML Research Intern

Deep Forest Sciences

Machine learning for drug discovery and materials science at a Y-Combinator backed startup.

• Shipped the image-preprocessing pipeline for the Prithvi drug discovery platform — automated object counting and UNet segmentation, deployed to production for client use
• Designed an LLM-driven generator for novel low-k dielectric materials with MD + DFT validation in the loop; bottleneck was 30+ day wet-validation cycles per candidate — cut to 5 days by pre-filtering on simulation, 83% reduction

Open Source Contributor & GSoC Mentor

DeepChem

Ship infrastructure that other researchers run experiments on. DeepChem (40k+ users) — open-source library for deep learning in drug discovery, materials, quantum chemistry, biology.

• Mentored GSoC 2024 — Target Conditioned Antibody Sequence Generation using Protein Language Models; guided contributor from spec → merged code over 12 weeks
• Shipped 7+ models — SCScore, UNet, OneFormer, ESM-2, ProtBERT — closing gaps in DeepChem's synthetic-feasibility, segmentation, and protein analysis stacks
• Designed GPU-accelerated parallel ODE solver infrastructure (PyTorch); bottleneck was sequential per-trajectory integration, replaced with batched parallel solve
• Optimized the image-loading pipeline and authored microscopy / cell-segmentation tutorials — drove a 10% site-traffic uplift

Visiting Research Fellow

Brown University — Serre Lab

Building grounded embodied agents that can reason about visual perspective and line-of-sight — a faculty children develop by age 1–2 and modern vision models still lack.

• Shipped VPTnav — Isaac Lab data pipeline + 20k-environment benchmark for visual perspective-taking; bottleneck was sim throughput, engineered batched collection to cut wall-clock 4 days → 8 hours (12×) and memory 6×
• Extended Isaac Lab's low-level sim interface to support runtime object re-coloring and scaling (not supported natively) — required for reason-balanced data generation
• Designed VPTnav around strict kinematic motion (no teleport shortcuts) and visual-variance-only object handling, forcing models to ground reasoning in observed geometry instead of training-set priors
• Built NavJEPA — action-conditioned latent-space predictor (ViT-S DINOv2 backbone, SIGReg, scheduled sampling) reaching val/cosine 0.677 on the dreamed-final-latent probe

VPTnav & NavJEPA

Synthetic benchmark and pipeline (Isaac Lab) for visual perspective-taking + action-conditioned JEPA predictor running mental simulation in latent space. Tests whether vision models can learn line-of-sight reasoning that human toddlers develop by 1–2 years old.

• PyTorch
• Isaac Lab
• JEPA
• Self-Supervised Learning
• Embodied AI
• VPT

mindweather — SAE Steering on Gemma 3

Mechanistic interpretability tinkering: use Gemma Scope 2 sparse autoencoders to find emotion-specific features in Gemma 3's residual stream, then steer generation by adding feature directions back at chosen scales. Multi-emotion mixing with signed scales.

• PyTorch
• Sparse Autoencoders
• Mechanistic Interpretability
• Gemma
• Steering

LMLF — LLM-Driven Molecule & Materials Generation

Semi-automatic pipeline using LLMs (GPT-4, Claude) with logical feedback loops for molecule generation, retrosynthesis, and low-k dielectric materials screening. Generated 20+ JAK2/DBH inhibitor candidates; screened 40+ dielectric materials with MD + DFT validation.

• Python
• LLMs
• RDKit
• Molecular Dynamics
• DFT
• Drug Discovery

6-DOF Robotic Arm + RL Digital Twin

Designed and built a 6-DOF robotic arm (servos, stepper motors, ESP32) with digital twin simulation for RL-based object localization and grasping using image and sensor data.

• Python
• PyTorch
• ESP32
• OpenCV
• Reinforcement Learning
• Robotics

Mixture of Experts for Named Entity Recognition

Implemented MoE layer from scratch and integrated into BiLSTM model for CoNLL 2003 NER — 12% accuracy increase and 32% F1-score improvement over baseline. Top-2 gating with load-balancing auxiliary loss.

• Python
• PyTorch
• HuggingFace
• Mixture of Experts
• NER

Time-Masked Autoencoders for Fluid Dynamics

Collaborated with Imperial College London on temporal masking in video autoencoders for fluid dynamics — predicting up to 20 future frames of Shallow Water simulations with 80% input masking, maintaining 80%+ SSIM.

• Python
• TensorFlow
• PyTorch
• Video Autoencoders
• Fluid Dynamics

Autonomous Rock Detection — Mars Rover

Led computer vision system for autonomous life detection. Transfer learning multi-class rock classifier for biological significance assessment. Multidisciplinary team of 7.

• Python
• TensorFlow
• Computer Vision
• Transfer Learning
• Robotics

Compact and Efficient RNA Representations with RNAvec Enable Residue-Level Interaction Mapping

Under review

Novel representation learning approach for residue-level RNA-protein interaction prediction using only primary sequences. Authors: Aaron R. Menezes, Omkar S. Sathe, Sanket R. Gupte, Aman A. Kattuparambil, Ashwin Srinivasan, Raviprasad Aduri (Aaron R. Menezes and Omkar S. Sathe contributed equally)

Open Source Infrastructure for Automatic Cell Segmentation

arXiv preprint

Open-source UNet and OneFormer-based segmentation models built into DeepChem with optimized image loading and tutorials. Authors: Aaron R. Menezes, Bharath Ramsundar

Open Source Differentiable ODE Solving Infrastructure

AAAI Workshop on AI to Accelerate Science and Engineering (AI2ASE)

GPU-accelerated parallel ODE solver infrastructure integrated into DeepChem for scalable neural ODE training. Authors: Rakshit Singh, Aaron R. Menezes, Rida Irfan, Bharath Ramsundar

Google Summer of Code (GSoC) Mentor

Google

Mentoring at DeepChem on Target Conditioned Antibody Sequence Generation using Protein Language Models

Excellence Award — International Rover Challenge

International Rover Challenge (IRC), India

Won Excellence Award for autonomous rock detection and biological significance classification system

2nd Place — Anatolian Rover Challenge

Anatolian Rover Challenge (ARC), Turkey

Led multidisciplinary team of 7 in designing rover for autonomous life detection in extreme terrain

INSPIRE Scholar

Department of Science and Technology, Government of India

Awarded to top 1% of science students in Class 12 HSC board exams; scholarship grant to support further studies in science