Ayan Mitra

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 — 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.

• 📄 Publications: arXiv listing · ADS library
• 💼 LinkedIn
• ✉️ ayan@illinois.edu
• 🔭 Currently: LSST Time-Domain Pipeline (NCSA / DESC), open to applied-ML / LLM / ML-engineering roles

---

🏆 Competitive Machine Learning & ENS Data Challenges

Project & Repository	Competition & Domain	Leaderboard Metric / Ranking	Official ENS Leaderboard Link	Tech Stack
📑 AMF_table_of_contents_extraction	ENS Challenge 86: AMF Financial PDF Document AI & ToC Section Extraction	🥉 Rank 3 in the World 0.4236 Public Score	View ENS Leaderboard	Python, LightGBM, scikit-learn, NLP
👁️ Valeo_industrial_computer_vision	ENS Challenge 157: Valeo High-Res Industrial Defect Computer Vision	🥈 Rank 2 Peak Validation 99.21% Accuracy	View ENS Leaderboard	PyTorch, torchvision, Pillow
💼 HrFlow_job_recommendation	ENS Challenge 164: HrFlow.ai Behavioral Job Recommendation & NLP	🏅 Rank 15 in the World 0.3144 Public Score	View ENS Leaderboard	PyTorch, Transformers, scikit-learn
🚆 SNCF_project	ENS Challenge 166: SNCF Transilien Platform Waiting Time Forecast	🚆 Rank 4 Local Benchmark 0.6370 MAE	View ENS Leaderboard	LightGBM, XGBoost, CatBoost, Optuna

---

🌌 Scientific ML, LLM AI Assistants & Open-Source Research

Project & Repository	Category & Description	Publication / Metric	Stack & Resources
🌌 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
🤖 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
🌌 Firecrown_wrapper_TD	Scientific Software: Open-source Python pipeline orchestrating Firecrown + CosmoSIS for supernova time-domain cosmology inference.	MIT Open-Source	Python, pytest, SACC

---

Major open-source contribution — SNANA

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 — 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	—	#1

Representative contributions:

• Generalized spline-interpolation library, wired into photo-z quantile and log-mass-vs-redshift estimation — #1664, #1669
• Data-driven host-galaxy weighting (WGTMAP) modes for simulation realism — #1632
• DiffSky host-galaxy catalog integration pipeline (magnitude joins, dedup, HDF5→pandas conversion) — #1640, #1704, #1728
• Core-collapse-supernova contamination prior implementation — #1524, #1534

Full contribution history → · Contributor graph →

---

Selected publications (ML-relevant)

• arXiv:2310.15649 (first author) — 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) — 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 — 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 — Simulation-based inference with conditional normalising flows + hierarchical Bayesian modelling to correct survey-selection effects in supernova cosmology (JAX / NumPyro).

(Full list: arXiv search)

---

Core skills

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#