~/work — projects · publications · certifications

Selected
Work

A snapshot of research projects, peer-reviewed papers, a patent, and ongoing learning.

Projects

Selected Builds

LLM · M.Tech Thesis 2024 — Present

Catalytic Depolymerization of Polyethylene — LLM-Curated Dataset & ML Yield Prediction

M.Tech thesis at IIT Madras under Prof. Jithin John Varghese & Prof. Tarak Patra.
End-to-end AI-driven literature mining pipeline using a local LLM to automate extraction of catalytic pyrolysis data, paired with ML models for hydrocarbon yield prediction.

Built an end-to-end AI-driven literature mining pipeline using Gemma4:e4b (Local LLM) to extract catalytic pyrolysis data from 736+ scientific publications
Curated a structured dataset of 559 experimental datapoints from 73 research papers — covering 120+ catalysts, 6 reactor configurations, and operating conditions for catalytic polyethylene pyrolysis
Built and optimized ML models (XGBoost, Random Forest, SVR) using Optuna-TPE for hydrocarbon product yield prediction — R² up to 0.78 for gas yield estimation
Engineered a fully local GPU-accelerated scientific workflow integrating OCR, PDF parsing, prompt engineering, and schema validation — drastically reducing manual data curation effort

Gemma (Local LLM)XGBoostRandom ForestSVROptuna-TPEOCR / PDF ParsingPrompt Engineering

Link coming soon

0.78

R² · Gas Yield

Scientific ML Jan 2025 — May 2025

Prediction of Absorption Energies of Chemical Species on Metal Catalyst Surfaces using ML Models

ML models to predict adsorption energies across 8 transition metals, replacing costly DFT calculations.

Developed and benchmarked 5 ML models to predict adsorption energies across 8 transition metals, achieving 0.12 eV MAE, reducing computational cost of DFT-based catalyst screening by >80%.
Engineered and automated descriptor discovery via PCA + varimax rotation, identifying 3 optimal physical descriptors that captured 98% dataset variance, outperforming conventional trial-and-error approaches.

XGBoostSVRPCAASEDFT

Link coming soon

0.12

eV MAE

Scientific ML Aug 2024 — Dec 2024

Optimization of Steam Methane Reforming Reaction to maximize conversion using Neural Networks

Deep NN to model methane conversion, enabling reliable process simulations without costly experiments.

Engineered and trained a deep neural network (8-layer, 50-neuron) to model methane conversion, achieving <0.5% prediction error versus numerical solvers, enabling reliable process simulations without costly experiments.
Automated reaction optimization using grid search over 100,000 parameter combinations, identifying optimal conditions that maximized CH₄ conversion by 71% while significantly reducing manual iteration time.
Optimized model architecture through systematic hyperparameter tuning, reducing weights by 80% with negligible loss in accuracy, demonstrating scalable deployment for data-driven process optimization.

PyTorchGrid SearchODEKeras

Link coming soon

71%

CH₄ Gain

Data Analysis · EDA 2025

Data-Driven Analysis of Friction Factor & Pressure Drop in Packed Bed Reactors

Multi-source packed bed dataset with end-to-end exploratory analysis and correlation modelling to identify dominant drivers of friction factor and pressure drop behavior.

Curated and analyzed a multi-source packed bed dataset of 1,549 experimental records spanning 11 references and 8 particle geometries; performed end-to-end EDA including missing value assessment, statistical profiling, and outlier detection to establish data quality baselines for downstream modelling
Engineered a correlation analysis pipeline combining Pearson correlation matrices, one-way ANOVA with η² effect sizes, and wall-effect subset analysis (D/d_p < 10) to quantify feature dependencies — identifying particle sphericity (η² = 0.995) and fluid type (η² = 0.96 for viscosity) as the dominant drivers of packed bed behavior

PandasNumPySciPyANOVAPearson CorrelationEDAMatplotlib

Link coming soon

1,549

Records · 11 Refs

Cheminformatics Jan 2025 — Apr 2025

De Novo Molecular Generation using AI-based Bigram Models

Developed a statistical AI language model (Bigram) to autonomously generate chemically realistic molecules.

Trained the model on a dataset of 3,000 SMILES strings, engineering a sampling process based on learned character transition probabilities to capture local molecular syntax and successfully generate 100 novel SMILES sequences.
Evaluated the structural uniqueness and chemical viability of the generated molecules by calculating Tanimoto Similarity coefficients with Morgan fingerprints via RDKit.
Achieved an average Tanimoto similarity score of 0.024 against the training dataset, demonstrating the generation of highly unique and distinct chemical structures.

RDKitSMILESBigramMorgan Fingerprints

Link coming soon

100

Novel Molecules

Engineering Aug 2022 — Dec 2022

Process Simulation of Ethyl Lactate - A Green Solvent

Simulated and optimized ethyl lactate production via reactive distillation using DWSIM with NRTL property model.

Simulated and optimized ethyl lactate production via reactive distillation using DWSIM, achieving 96.13% lactic acid conversion & 95.7% ethyl lactate yield.
Designed a process flowsheet integrating fermentation, reactive distillation, and product purification, resulting in an ethyl lactate production rate of 234 g/h and energy consumption of 111.41 kJ/g.
Applied the NRTL property to represent non-ideal mixtures and optimize operating parameters for high-purity.

DWSIMNRTLReactive Distillation

Link coming soon

96%

Conversion

Publications & Patents