Applied machine-learning scientist and research software engineer. I build and evaluate state-of-the-art Large Language Model (LLM) architectures, NLP pipelines, Document AI / Layout ML models, high-resolution Computer Vision, and Time-Series classifiers β with a focus on distribution shift, uncertainty quantification, and reproducible production pipelines.
Background: PhD in astrophysics (supernova cosmology, LSST/DESC pipelines). Currently applying that research experience to general applied-ML problems: LLM & AI assistants, document structure extraction, competitive computer vision, time-series forecasting, and probabilistic ML.
- π Publications: arXiv listing Β· ADS library
- πΌ LinkedIn
- βοΈ ayan@illinois.edu
- π Currently: LSST Time-Domain Pipeline (NCSA / DESC), open to applied-ML / LLM / ML-engineering roles
| Project & Repository | Competition & Domain | Leaderboard Metric / Ranking | Official ENS Leaderboard Link | Tech Stack |
|---|---|---|---|---|
| π AMF_table_of_contents_extraction | ENS Challenge 86: AMF Financial PDF Document AI & ToC Section Extraction | π₯ Rank 3 in the World0.4236 Public Score |
View ENS Leaderboard | Python, LightGBM, scikit-learn, NLP |
| ποΈ Valeo_industrial_computer_vision | ENS Challenge 157: Valeo High-Res Industrial Defect Computer Vision | π₯ Rank 2 Peak Validation99.21% Accuracy |
View ENS Leaderboard | PyTorch, torchvision, Pillow |
| πΌ HrFlow_job_recommendation | ENS Challenge 164: HrFlow.ai Behavioral Job Recommendation & NLP | π
Rank 15 in the World0.3144 Public Score |
View ENS Leaderboard | PyTorch, Transformers, scikit-learn |
| π SNCF_project | ENS Challenge 166: SNCF Transilien Platform Waiting Time Forecast | π Rank 4 Local Benchmark0.6370 MAE |
View ENS Leaderboard | LightGBM, XGBoost, CatBoost, Optuna |
| Project & Repository | Category & Description | Publication / Metric | Stack & Resources |
|---|---|---|---|
| π GWCCSN_EOS_Ye | Scientific ML: 1D-CNN classifying nuclear equation-of-state from core-collapse supernova gravitational-wave time series. | 86.5% Held-out Acc (arXiv:2310.15649) |
TensorFlow, scikit-learn, Gradio Live Demo |
| π€ SNANA_PIPELINE_ASSISTANT | LLM & Agentic AI Assistant: Autonomous LLM agent for configuring, validating, and executing complex SNANA time-series analysis workflows. | Production AI Assistant | Python, OpenAI/Gemini LLM APIs, LangChain |
| π Firecrown_wrapper_TD | Scientific Software: Open-source Python pipeline orchestrating Firecrown + CosmoSIS for supernova time-domain cosmology inference. | MIT Open-Source | Python, pytest, SACC |
SNANA is the supernova simulation, light-curve fitting, and time-domain analysis package used across the Rubin Observatory LSST Dark Energy Science Collaboration and prior DES/SDSS-era SN cosmology surveys β facility software, not a personal project, used and validated by a large external collaboration. I'm its 2nd-largest contributor by code volume:
| Lines changed (all-time) | Merged PRs | Rank | |
|---|---|---|---|
| am610 | 181,032 | 74 | #2 of 28 contributors |
| Lead maintainer (RickKessler) | 2,693,737 | β | #1 |
Representative contributions:
- Generalized spline-interpolation library, wired into photo-z quantile and log-mass-vs-redshift estimation β #1664, #1669
- Data-driven host-galaxy weighting (
WGTMAP) modes for simulation realism β #1632 - DiffSky host-galaxy catalog integration pipeline (magnitude joins, dedup, HDF5βpandas conversion) β #1640, #1704, #1728
- Core-collapse-supernova contamination prior implementation β #1524, #1534
Full contribution history β Β· Contributor graph β
- arXiv:2310.15649 (first author) β 1D CNN classification of nuclear equation-of-state from core-collapse-supernova gravitational-wave time series; robustness across sampling rate, signal window, and physical nuisance parameters.
- arXiv:2402.18124 (first author) β Neural regression with genetic-algorithm hyperparameter search (nnogada) and Monte Carlo dropout for uncertainty-aware reconstruction of cosmological observables from simulated Rubin/LSST data.
- arXiv:2409.14508 β Benchmarked CNNs, RNNs, and six classical ML methods (random forest, SVM, XGBoost, etc.) on gravitational-wave time-series classification; quantified performance drop under simulation-domain mismatch.
- arXiv:2603.11165 β Simulation-based inference with conditional normalising flows + hierarchical Bayesian modelling to correct survey-selection effects in supernova cosmology (JAX / NumPyro).
(Full list: arXiv search)
LLM & NLP: Large Language Models (LLMs), Sentence-Transformers, Document AI / Layout ML, Hugging Face Transformers, Agentic AI Assistants, Prompt Engineering, Text Embeddings ML & Computer Vision: PyTorch, LightGBM, XGBoost, CatBoost, Computer Vision (ConvNeXt, ResNet, 4-Way TTA), TensorFlow, scikit-learn, uncertainty quantification, simulation-based inference, time-series forecasting Scientific computing: JAX, NumPyro, HPC / batch pipelines, SNANA, CosmoSIS, Firecrown Engineering: Python, Git, Docker, pytest, LaTeX, R, C#