Headline

Data Scientist | ML Engineer | Deep Learning · Computer Vision · Physics-Informed ML | PyTorch · Azure · AWS | IISc

About

My path into ML started at IISc, where my thesis on robust object detection pulled me toward a specific kind of problem: models that have to hold up when real-world conditions stop cooperating. At Mercedes-Benz R&D, I built physics-informed deep learning systems for EV battery thermal prediction, integrating thermodynamic constraints directly into loss functions so models respected the physics, not just the data. I also worked on VAE-based anomaly detection for battery lifecycle monitoring across vehicle fleets, and LLM-powered document automation using RAG and LangChain. Co-inventor on 2 filed global patents from this work. At Picarro, I build ML systems where being wrong has real consequences. I replaced a 15-second classical solver with a 10ms deep learning inference pipeline for real-time spectroscopic gas analysis, and built an atmospheric source inversion system using Bayesian MCMC over physics-based dispersion models to localize industrial emissions with quantified uncertainty. My IISc thesis focused on robust object detection under real-world visual corruptions — improving Mask R-CNN performance under distribution shift through custom augmentation and regularization. Published at ETAAV 2025. 2 filed global patents. 1 published paper. M.Tech, IISc Bangalore. Looking for roles where research depth and the discipline to ship both matter.

Location

Bengaluru

AI stack

claude

cursor

langchain

mcp

notebooklm

ollama

Work experience

Data Scientist at Picarro, Aug 2025 - Present · 1 yr 1 mo, Bengaluru, Karnataka, India, Building deep learning systems for real-time spectroscopic gas analysis and industrial atmospheric monitoring — from model architecture through production deployment on AWS. Work spans physics-informed deep learning, Bayesian inference, distributed GPU training, and MLOps. • Replaced a 15-second classical optimization model with a 10ms deep learning inference pipeline for real-time gas concentration measurement; improved compound detection accuracy from 78% to 84% using a 1D Vision Transformer with compound query cross-attention on infrared spectroscopic data. • Generated 200M synthetic spectral mixtures via physics-based simulation and trained across 4 GPUs (AWS g4dn.12xlarge), cutting training time from 26 hours to 2.5 hours; validated model performance against real-world sensor data. • Built an atmospheric source inversion system for industrial fenceline monitoring — applied Bayesian MCMC over physics-based dispersion models to localize emission sources and quantify uncertainty from continuous multi-sensor data. • Designed a multi-source emission identification approach using sparse regularization on a physics-based forward model dictionary, recovering active source locations among an unknown number of simultaneous emitters without requiring prior knowledge of source count. • Established MLflow-based experiment tracking and model monitoring with KPIs covering prediction drift, data quality, and production model performance. • Present model performance and business impact in quarterly business reviews to technical and executive stakeholders.

Junior Data Scientist at Mercedes-Benz Research and Development India, Aug 2023 - Jul 2025 · 2 yrs, Bengaluru, Karnataka, India, Built end-to-end ML pipelines on Azure for EV battery systems — spanning physics-informed model training, anomaly detection, generative AI, and production API deployment. Co-inventor on 2 filed global patents from this work. • Designed and deployed a physics-informed deep learning model for EV battery thermal prediction — integrated thermodynamic constraints directly into loss functions to improve physical consistency across operating conditions, trained on synthetic data and validated against real vehicle sensor readings. • Developed VAE-based anomaly detection for battery lifecycle monitoring across multiple vehicle models, enabling early detection of degradation patterns from fleet telemetry. • Built a RAG pipeline using LangChain and GPT-based models with FAISS vector search for automated document summarization, reducing manual review from hours to minutes. • Developed a FastAPI backend for real-time EV route optimization — integrated charging station data from PostgreSQL, implemented round-robin load balancing across Azure Functions, and served fleet-level inference requests. • Deployed models to production via MLflow and Azure ML with CI/CD pipelines, collaborating with product managers and software engineers across the full development cycle.

M.Tech Researcher at Indian Institute of Science (IISc), Aug 2021 - Jul 2023 · 2 yrs, Bangalore Urban, Karnataka, India, Thesis research at IISc's Artificial Intelligence and Robotics Laboratory on domain generalization for instance segmentation in aerial imagery — building deep learning models resilient to the sharp performance drops caused by real-world shifts in lighting, weather, and atmospheric conditions. • Developed Metacognitive Mask R-CNN in PyTorch — integrated a self-adaptive binary loss selection mechanism and SpatialDropout2D regularization into a ResNet-101 + FPN backbone, enabling robust learning from just 455 annotated aerial images. Achieved mAP 69.3 on NWPU VHR-10, outperforming CPISNet (67.5), ARE-Net (64.8), and Precise Mask R-CNN (64.8). • Systematically evaluated five augmentation strategies (APR, GridMask, Random Erasing, Gaussian Blur, Color Jitter) and three dropout variants (SpatialDropout2D, DropBlock, Standard Dropout) through full ablation studies — SpatialDropout2D combined with APR delivered the strongest robustness, improving detection by 9.5% and segmentation by 6.5% under extreme brightness corruption versus baseline. • Published at IEEE ETAAV 2025 — equal first-author contribution

Education

Master of Technology - MTech, Aerospace Engineering, Indian Institute of Science (IISc)

Bachelor of Technology - BTech, Electrical Engineering, BIT Sindri

Public links

github: https://github.com/sayanchaudhuri14

leetcode: https://leetcode.com/sayanchaudhuri