Swagat Padhan
Machine Learning Engineer & Researcher
AI Engineer and TSMC Graduate Fellow specializing in Embodied AI, Bayesian Machine Learning, and Large Vision-Language Models. Building production-grade AI systems from research to deployment.
Experience
Over 3 years building production AI systems across industrial automation, health-tech, and academic research.
Graduate Research Fellow
Logos Lab, Arizona State University
May 2024 – Present
- •Developed Metric-Semantic-Predicate (MSP) kernels using Variational Inference in Pyro/PyTorch, achieving 31% improvement in grounding accuracy
- •Engineered Product-of-Experts fusion mechanism combining BNN likelihoods with VLM priors, reducing metric error by 42%
- •Built custom data collection pipeline in Habitat-Sim with real-time probability heatmaps (48 scenes, 3,400+ spatial relations)
- •Implemented rigorous evaluation using Wasserstein Distance with 1,000 Monte Carlo samples per query
Machine Learning Engineer Intern
Hobbes Health, Seattle, WA
May 2025 – Jul 2025
- •Architected agentic RAG meal-planning engine using Gemini 2.0 Flash with async Chain-of-Thought reasoning
- •Deployed scalable microservice on GCP Cloud Run with FastAPI, Cloud SQL, and JSON Schema validation
- •Engineered semantic caching layer using SBERT embeddings, reducing API calls by 38%
- •Achieved 16% increase in plan acceptance rates and 12% increase in 7-day user adherence via A/B testing
Applied AI Engineer
Birla Carbon (Aditya Birla Group), India
May 2019 – Jul 2022
- •Deployed real-time anomaly detection across 30+ manufacturing plants using FFT and Isolation Forests
- •Prevented 24 critical failures annually, reducing MTTR by 42%
- •Developed RESTful inference APIs on AWS EC2, achieving verified annual savings of $240,000
- •Led digital transformation as MAGNET Ambassador, implementing Docker and CI/CD pipelines
Projects
Research and production projects spanning AI, Robotics, and Systems Programming.
Enhanced Agentic RAG Pipeline for Financial Analysis
Multi-modal knowledge base processing 7 SEC filings with specialized agent tools including RAG Librarian, SQL Analyst, Trend Analyst, and Web Scout. Advanced reasoning engine with Gatekeeper, Planner, and Auditor for self-improvement.
- →371 structured chunks with 1,660 vector embeddings
- →CrossEncoder re-ranking for query optimization
- →State graph with conditional routing and self-correction
GraphEQA Sandbox for Spatial Reasoning
Extended state-of-the-art GraphEQA framework for metric-spatial queries in cluttered indoor scenes. Built deterministic sense-plan-act loop with RGB-D perception and online metric-semantic scene graphs.
- →Real-time object state tracking
- →Dynamic spatial relationship modeling
- →Metric-semantic integration
IntuitionAI: Adaptive Learning Platform
Full-stack adaptive learning platform with RAG-based multi-agent system. Integrated GPT-4 and Pinecone for real-time personalized content delivery across 12+ content types.
- →Multi-modal transformer analysis
- →Reinforcement learning for curriculum optimization
- →40% boost in learning efficiency
Offline Reinforcement Learning for Locomotion
Replicated Behavior Proximal Policy Optimization (BPPO) algorithm on D4RL Hopper-Medium-v2 dataset. Validated offline monotonic policy improvement with CUDA optimization.
- →Hyperparameter optimization (clip ratio, decay rate)
- →NVIDIA RTX 3060 acceleration
- →Improved stability across noisy datasets
MiniBase Disk Manager Enhancement
Extended MiniBase database system with columnar storage engine. Integrated Bitmap Indexes and B+ Trees, achieving 36% improvement in columnar query performance.
- →Columnar storage with predicate pushdown
- →Custom buffer manager redesign
- →Complex B+ Tree deletion and rebalancing
Inverted Pendulum System
Designed, fabricated, and controlled an inverted pendulum system with variable load handling on edge devices. Undergraduate thesis project demonstrating control theory and embedded systems.
- →Variable load handling capability
- →Edge device implementation
- →Real-time control system
Research
Graduate research focused on probabilistic models, embodied AI, and spatial reasoning.
Probabilistic Metric-Semantic Grounding for Embodied AI
Composing Bayesian Spatial Kernels, VLMs, and 3D Scene Graphs
Committee
Dr. Nakul Gopalan (Chair), Dr. Chitta Baral, Dr. Hani Ben Amor
Key Contributions
- •Novel Metric-Semantic-Predicate kernels using Variational Inference achieving 31% accuracy improvement
- •Product-of-Experts fusion of BNNs and VLMs reducing metric error by 42%
- •Human-in-the-loop dataset with 48 scenes and 3,400+ spatial relations
Research Interests
Probabilistic Machine Learning
Bayesian Neural Networks, Variational Inference, Gaussian Processes, Uncertainty Quantification
Embodied AI & Robotics
Spatial Reasoning, 3D Scene Understanding, Robot Navigation, Grounded Language Models
Large Language Models
Vision-Language Models, Agentic Systems, RAG Architectures, LLM Optimization
Production ML Systems
Model Deployment, API Design, Real-time Inference, Cloud Infrastructure, MLOps
Publications
Thesis publication in progress. Check back for updates on peer-reviewed publications.
Let's Connect
Interested in collaboration or have questions? Feel free to reach out.
Connect Online
Available for full-time opportunities starting October 2025
Open to discussing opportunities in ML Engineering, AI Research, and Robotics.