Rigor first
An MIT PhD designing molecules with graph neural networks, published in Nature Chemistry and ACS Central Science. I still read the paper before I buy the platform.
Director of Analytics · Caterpillar Inc.
I build the AI that runs physical industry.
I lead a 90+ person AI, data science and product organization putting digital twins and agentic AI to work across Caterpillar's global manufacturing. Before that: an MIT machine-learning PhD, a startup I co-founded, and strategy for a $15B segment. I still write code.
Answers are retrieved from this site, not generated: a small embedding model runs in your browser, and nothing you type leaves your device.
90+people in the organization I lead
150+facilities in platform scope
15+peer-reviewed papers
$30Mannual platform budget
CaterpillarMITGoogleAmgenNVIDIA GTC · CESTEDx
Hands on the keyboard
I prototype the digital twin assistant, the RAG pipeline and, yes, this site's in-browser model. Credibility with engineers is earned in code reviews.
Value at scale
Founder with a P&L, chief of staff in a $15B segment, and now an AI leader accountable for outcomes across 150+ facilities.
Trajectory
One arc: optimizing physical reality with computation.
The model that predicted failures in a peptide synthesizer at MIT is a cousin of the one predicting failures on a Caterpillar line. Each role built a reusable asset the next one used. At Caterpillar that meant Manager to Director in under three years, with two promotions in 2025, across procurement, mining, strategy and manufacturing.
-
- Lead a 90+ person organization as a leader of leaders: managers of data scientists, AI/GenAI engineers, simulation developers and product managers across the US and India, plus a senior principal architect.
- Own the AI and product roadmap for the Modern Manufacturing digital platform: predictive maintenance, inventory, dynamic scheduling, connected assets and operational metrics.
- Architected a conversational assistant over a 3D digital twin on CUDA-accelerated NVIDIA AI Factory, Omniverse and NeMo agents; piloted across four US and UK plants.
- Two invention disclosures: hierarchical clustering for predictive maintenance in low-data regimes, and hybrid RAG for supply chain.
- Promoted to Director in 2025, my second promotion that year.
$30MAnnual platform budgetGTC · CESWork shown at NVIDIA GTC DC 2025 and CES 20264 → 150+Pilot plants to global rollout plan -
- Promoted to Senior Manager to run strategy for the segment.
- Led the enterprise strategy workstream for sustainable mining, reporting into the Group President's office.
- Built the TAM/SAM framework with AI-driven adoption modelling across diesel, electric and hybrid propulsion.
- Informed M&A and venture decisions in autonomy and digital mine software; prepared board-level and Investor Day material.
$3B+Market opportunity identifiedBoardLevel strategy and capital-allocation work -
- Architected an OEM-agnostic data platform linking 10+ subsystems: battery trucks, fast chargers, energy storage, fleet and energy management.
- Negotiated interoperability and encryption protocols with OEM partners, enabling commercial data services.
- Shipped Early Learner dashboards with agentic workflows and natural-language query for the top five global miners.
10+Subsystems on one data platform60%Less analyst reporting time -
- Co-led the enterprise supply resiliency digital thread: scenario planning and risk across 1.5M parts, 7,000+ suppliers and $20B of annual spend.
- Designed ChatDocs, a GenAI platform (LLMs, hybrid RAG, fine-tuning) that extracts terms from contracts, quality certificates and risk reports.
- Led 12 Purdue Data Mine students building predictive risk analytics over 110+ datasets for 180 countries.
$1M+Annual savings from ChatDocsCEO AwardTop 1% of 120K+ employees -
- Built graph neural networks that predict reaction impurities in small-molecule drug synthesis, and an impurity propagation framework across multi-step routes.
- Adapted inverse structure elucidation from mass spec and NMR, cutting unknown-compound identification from weeks to hours.
- Presented at NeurIPS 2022 (AI for Accelerated Materials Design workshop) within four months; became my MIT Sloan thesis.
$120K+Saved per unknown identificationWeeks → hoursStructure elucidation time -
- Advised by Prof. Rafael Gómez-Bombarelli. Thesis: Designing Macromolecules Using Machine Learning and Simulations.
- Closed AI → synthesis → experiment loops with 30+ researchers across five labs, raising experimental throughput 10x.
- Open-sourced GLAMOUR and Peptimizer; helped secure $2.27M in grants; mentored nine students.
- TEDxBoston speaker on how AI can make rare diseases rarer.
15+Papers · Nature Chemistry, ACS Central Science1US patent application (AI-designed delivery peptides) -
- Built ML models for CRISPRi guide RNA design across all 20,000 human genes, beating the prior state of the art by 20%+.
- Co-developed tooling for automated model building, hyperparameter tuning, architecture search and feature attribution.
20%+Accuracy gain over prior SOTA -
- Scaled from two founders to a 30-person team across the US and India, with full P&L ownership.
- Ran the sales cycle end to end and built the customer success function.
- Flagship RPA/ML product automated customer-inquiry processing for an oil and gas pipeline client.
30Person team built from zero$3M+Annual client savings -
- Institute Gold Medal (rank 1 of 110); finished a four-year degree in three.
- SN Bose Scholar at Harvard (cancer immunotherapy) and OIST fellow (computational genomics).
- Quantitative researcher at EIS Global, Singapore: 20%+ lift in high-frequency trading models.
- Technical Executive at UIDAI (Aadhaar), delivering $3M analytics programs and privacy policy analysis.
Rank 1Of 110, IIT Roorkee
Moments
Ten years, four rooms.
Capabilities
Four disciplines, one operating system.
Selected work
Built, shipped, measured.
Each project below is in production or shaped a capital decision. Numbers are the business outcome, not the model metric.
Recent builds
Still shipping code.
Each build starts from a question I keep running into at work and answers it on my own time with this year's tools: world models and OpenUSD for digital twins, small language models that run on a phone or in the browser, and teams of AI coding agents working in parallel. Everything is open on GitHub.
Spatial AIIn progress
Auto-Twin
Digital twins usually start with weeks of laser scanning and CAD. Can a walkthrough video and a world model stand one up for about twenty cents of compute?
A video tour of the Central Texas Food Bank becomes a Gaussian-splat world through the World Labs API, is converted to OpenUSD, bound to live telemetry on AWS and rendered in the browser, Blender or Omniverse. Cost guards make every paid call resumable and never double-billed.
Digital twinLive
Michelin Kitchen Twin
Factory planning ideas like MRP, routings and OEE are hard to show with plant data nobody can share. A famous kitchen makes them visible to anyone.
A 3D twin of a three-star kitchen running a full tasting-menu service. Covers are the build plan, 53 recipes are bills of materials and routings, and a discrete-event simulation plays out tickets, stations and bottlenecks minute by minute. Parallel coding agents built the product, data, simulation and 3D layers.
World modelsLive · 2 papers in draft
Baby Senses Lab
AI world models train on adult-quality data from day one. Infants start with blurry, muffled input and still learn fast. What does that staging buy a model?
One world model per sense (vision, hearing, touch, body position), each fed through an age-tuned filter and grown as it trains, then fused. The public site lets parents see and hear like a baby of any age; researchers get the findings, papers and code.
On-device AIv1.1 released
LabTrends
A family's lab results sit in PDFs and paper scans from different clinics, and seeing a trend usually means handing health data to someone's cloud.
An Android app that reads reports with on-device OCR, parses biomarkers with rules backed by a small local language model, and has a person confirm every value before it is saved. The release build has no internet permission at all.
Code ↗Updated Oct 2026
In-browser AIIn progress
Academic Matchmaker
Prospective PhD students choose advisors from rankings and cold emails rather than research fit.
Drop in a CV or paper and a small language model running in the browser on WebGPU extracts research interests; embeddings then rank faculty across the top 100 universities. Nothing leaves the device. A scheduled pipeline is rebuilding the faculty index from open scholarly data.
Code ↗Updated Oct 2026
Physical AILive
Digital Twin History
“Digital twin” means something different in every meeting. A shared, cited history gives teams a common vocabulary.
An open dataset of milestones from andon lights and Apollo's simulators to AI world models and agentic twins, with a walkable 3D factory shown across six eras and research trends from OpenAlex. MIT licensed for courses and talks.
AI and workTechnical paper
India Iceberg Index
Most AI-and-jobs estimates count tech jobs. The larger exposure sits in everyday tasks across the whole economy.
Adapts the Iceberg Index method to India: generative-AI exposure scores mapped to national occupation codes and weighted by 2023–24 labor survey data, district by district, with an interactive dashboard and a technical paper.
Code ↗Updated Dec 2025
Playbook
How I think about data and AI.
A point of view across strategy, data foundations, context, the AI platform, agents, governance, economics, organization, adoption, how programs mature and how to measure them. Ten beliefs, twelve chapters, written to be argued with.
Recognition
On stage and on the record.
Writing
Writing, weekly.
Published on LinkedIn to 23,000+ followers, from new models and agents to AI on the factory floor. Browse all essays.
FAQ
Short answers.
Who is Somesh Mohapatra?
Somesh Mohapatra, PhD, MBA, is an AI, data science and product leader. He is Director of Analytics at Caterpillar Inc., where he leads a 90+ person organization building digital twins and agentic AI for global manufacturing. He holds a PhD in machine learning for molecular design from MIT and an MBA from MIT Sloan's Leaders for Global Operations program.
What does Somesh Mohapatra do at Caterpillar?
He is Director of Analytics and Head of Data Science and Product Management for Caterpillar's Modern Manufacturing digital platform, covering predictive maintenance, inventory, dynamic scheduling, connected assets and operational metrics across 150+ facilities. Earlier at Caterpillar he was chief of staff for strategy in the $15B Resource Industries segment, led minesite electrification analytics, and built ChatDocs, a GenAI platform for supplier contracts that earned the CEO Award.
How large are the teams and programs Somesh Mohapatra has led?
He leads a 90+ person organization as a leader of leaders, with managers of data scientists, AI and GenAI engineers, simulation developers and product managers across the US and India, and owns the AI and product roadmap for a manufacturing digital platform scoped to 150+ facilities, with a $30M annual budget. Earlier at Caterpillar he co-led a supply resiliency program covering 1.5 million parts, 7,000+ suppliers and $20B of annual spend, and did strategy and capital-allocation work for the $15B Resource Industries segment; he was promoted twice in 2025, to Senior Manager and then Director. As a startup co-founder he grew a team from two to 30 people with full P&L ownership.
Where did Somesh Mohapatra study?
He earned a PhD in machine learning for molecular design at MIT (2018 to 2022), advised by Prof. Rafael Gómez-Bombarelli, a dual-degree MBA from the MIT Sloan Leaders for Global Operations program, and a B.Tech from IIT Roorkee, where he won the Institute Gold Medal as rank 1 of 110.
What is Somesh Mohapatra's research about?
His research applies machine learning to molecules, materials and manufacturing: graph representations of macromolecules (GLAMOUR), deep learning for peptide design and fast-flow peptide synthesis, and reaction impurity prediction for drug manufacturing. He has 15+ peer-reviewed papers, including two in Nature Chemistry and several in ACS Central Science, and a US patent application on AI-designed drug delivery peptides.
How does Somesh Mohapatra approach enterprise AI?
His playbook treats data and AI as a multiplier for the business a company already wants to become: start from the business's core learning problem, fix data foundations before models, run one governed AI platform with risk-tiered governance, grant agents autonomy per workflow, and measure decisions changed and value created rather than pilots launched.
What does Somesh Mohapatra speak and write about?
He writes and speaks about enterprise AI, digital twins, agentic systems and AI in manufacturing, supply chain and operations, through his LinkedIn essay series, invited talks and his playbook on data and AI at scale. He is always glad to compare notes with practitioners, researchers and leaders working on the same problems.
What is Somesh Mohapatra building and writing about now?
He builds in the open on GitHub (github.com/pikulsomesh): browser-based digital twins, spatial AI that turns a walkthrough video into an OpenUSD twin, developmental world models and on-device AI. He writes weekly on LinkedIn about open-weight models, inference economics, agent design and evaluation, and what AI does to work, teams and judgment.
Where is Somesh Mohapatra based, and how can I contact him?
He is based in the Dallas–Fort Worth area of Texas and works across the US, UK and India. The best ways to reach him are email (from the contact section of someshmohapatra.in) or LinkedIn at linkedin.com/in/pikulsomesh.
Contact
Let's talk about
Let's talk about
AI at scale.
I am always glad to compare notes with people working on AI in manufacturing, supply chain and operations, and to speak or write about enterprise AI and digital twins.