Build notes first: each thing I build, written up in parts with the code, the architecture and what broke along the way. Then what I publish on LinkedIn, sorted by topic: new models and what they cost to run, agents and the software around them, physical AI in heavy industry, and what AI does to work, plus talks and conversations I have been part of.
A working digital twin of a representative Texas food bank, built from a public news video in seven days for about $1.54 of compute. No CAD drawings, no laser scans, no site visit. Five parts, from why to the full pipeline and the paper.
Latest: Part 5, The whole pipeline, what it cost, and the paper
How I built a private, in-browser search that matches a CV against 181,568 researchers at the world's top 200 universities. Covers fitting the index into a static site, semantic matching with a small model, and why concept tags had to come next.
Latest: Part 3, Semantic matching in the browser, and why it needed concept tags
A browser-based digital twin of The French Laundry, a three-Michelin-star kitchen, running a full dinner service, modeled the way a factory is. Recipes are bills of materials, covers are the build plan, and a deterministic simulation drives a photoreal 3D kitchen and an operations dashboard.
Latest: Part 5, How five parallel AI agent threads built the kitchen twin
Why I built an open dataset and interactive site tracing digital twins from 1880s time clocks and Apollo's simulators to live factory twins and AI world models, with 53 cited milestones, a walkable 3D factory in six eras, and research trends from OpenAlex.
A digital twin mixes measured, designed, published, estimated and simulated data. Saying which is which, on every record, is what makes the real parts trustworthy. A practical pattern from three builds.
When several AI coding agents build one system in parallel, the speed comes from clear ownership and a written contract between the parts, not from the agents. The same lesson that makes human teams fast, learned again across three builds.
I have always been fascinated by how kitchens work. A great restaurant on a busy night is one of the most demanding production systems there is: hundreds of plates, each one made to order, timed to the minute, and inspected before it…
For close to a year, I have been working on one idea from two directions at once. - Approach 1 - machine learning: how should we train models that learn about the world by predicting what they will sense next? - Approach 2 - infant…
Digital twins have been a passion project of mine for a long time, and the part that fascinates me most is the history. Background - All through my time at Massachusetts Institute of Technology / MIT Leaders for Global Operations (LGO) - I…
My family's lab reports live in a dozen places: PDFs in email, photos in chat threads, paper in a drawer, and even when you change PCPs - different applications for different health care providers. Every time a doctor asks "what was it…
System One Models or Contrastive Language Models have taken the feeds by storm with Jev by TypeSafe AI (super-cheap API), Julia-1 by Supersonic Labs (at 144M), and recently CLM-8B by Stanford University and NVIDIA(open source, 8B). What…
As autonomous coding agents become standard tooling across software engineering, the primary bottleneck in AI-assisted development has shifted from raw model capability to prompt and context structure. Simply passing massive, unstructured…
We often assume that generative AI will seamlessly remove the friction from the innovation process. However, recent research published in the Harvard Business Review reveals a more complex reality. AI frequently deepens the human…
The debate between proprietary APIs and open-weight architectures is reaching a critical inflection point. Just recently, I discussed the radical feasibility experiment of running Moonshot AI's 2.78-trillion-parameter Kimi (Moonshot AI) K3…
Just a month ago, the idea of having a 2.8T parameter model open-sourced, let alone hosting it was completely unthinkable, even with aggressive quantization. A recent open-source engineering exercise called kimi-k3-in-c by developer Fareed…
When predicting how artificial intelligence will change the workforce, most analysts begin with a fixed list of tasks associated with a specific job and ask which ones a model can automate, similar to my personal research inspired by the…
When the automobile was first introduced, early designs were literally called horseless carriages. Manufacturers took the existing form factor of a wooden buggy, removed the horse, and installed a small engine. It took years for designers…
The history of artificial intelligence is a pendulum swinging between absolute control and creative chaos. In the 1950s, early AI models were envisioned as massive, deterministic decision trees. Post 1990s came neural networks, plunging…
In the rush to integrate artificial intelligence into daily operations, we are confronting a dangerous psychological side effect: cognitive surrender. A recent article in the Harvard Business Review by Melchior Tamisier-Fayard, Theodoros…
For a long time, the prevailing wisdom in the artificial intelligence community was that true frontier models, those exceeding a trillion parameters, would remain locked behind proprietary APIs. The compute required to train and run them…
To the experienced ear, the clatter of a diesel engine or the hum of a turbine speaks volumes. It is akin to a doctor listening to a patient's heartbeat. On any given factory floor, you can find a veteran technician who will tilt his head,…
When engineering organizations rush to demonstrate return on investment for their artificial intelligence initiatives, the entry-level developer workforce is often the first target for cost reduction. The logic seems simple: automate…
As developer workflows shift toward agentic AI systems like Anthropic Claude Code and Cursor, we are encountering a new type of architectural bottleneck: command-line interface noise. When an autonomous agent executes terminal commands to…
Imagine an operations manager arriving at a job site early in the morning, reviewing the fleet metrics, and noticing something unusual. Overnight, without a single technician touching the equipment, every excavator and forklift in the yard…
As artificial intelligence systems transition from simple, single turn interactions toward complex, coordinated multiagent workflows, low latency inference becomes a primary engineering constraint. Standard autoregressive large language…
When we evaluate how artificial intelligence alters the enterprise, the conversation usually slides into a predictable binary: full automation versus basic productivity gains. At the recent AI and Society Forum at Massachusetts Institute…
In industrial robotics and the development of digital twins, a recurring challenge is the reality of mixed-vintage facilities. We frequently see state-of-the-art autonomous systems operating directly alongside legacy machinery. In these…
A common question echoing across enterprise teams right now is, "What 10% of my job should I focus on to maintain a competitive edge against AI? Marco Argenti, CIO at Goldman Sachs, recently addressed this in the Harvard Business Review,…
Gemma 4 12B by Google DeepMind / Google brings high-performance multimodal intelligence directly to consumer laptops. By bridging the gap between edge-friendly models and advanced large-scale architectures, it packages massive capabilities…
In the discussion surrounding advanced artificial intelligence, a critical question remains: are we building systems to automate humans out of the loop, or are we designing tools to amplify human capability? The core philosophy of the…
The startup landscape is undergoing its most profound transformation since the internet revolution. A Harvard Business Review article by Vivian S. Lee, MD, PhD, MBA, Linda Mantia, and Jon McNeill outlines exactly how agentic artificial…
When a multi-agent system fails, the root cause is rarely a single bad response. It is often a structural issue, such as a handoff occurring too late or a specialist agent missing a critical signal across multiple runs. To improve these…
For over a century, industrial mining has been defined by brute force. Massive rotary drills punch continuous holes into the earth, breaking ground based on geological surveys and core samples that might be weeks or months old. The…
The industrial sector often operates under the assumption that artificial intelligence is a finished product that management buys and installs to automate the frontline workforce. A recent piece in the Harvard Business Review by Tracey…
In developing agentic workflows, the hardest part is rarely the model's reasoning. The real friction is orchestration. Running multiple autonomous agents across different tasks usually turns into a chaotic management problem. Google just…
In the age of cloud computing, some of the most critical heavy equipment does not move earth or manufacture goods. It fights heat. Picture a sprawling data center on a sweltering day. Industrial chillers, massive HVAC systems, and heavy…
We spend a lot of time discussing modern cloud architectures, but the reality is that the backbone of global civilization still runs on mainframes. Financial services, government infrastructure, and manufacturing production controls rely…
For decades, enterprise software companies built their competitive moats around the user interface. Workforces developed muscle memory for navigating specific dashboards, and that operational friction kept companies locked into legacy…
In the heart of a power plant, a gas turbine the size of a bus spins at thousands of revolutions per minute, generating the electricity that entire cities depend on. The stakes are incredibly high. A single cracked turbine blade can…
Perplexity officially launched its Personal Computer app for Apple MacOS. This moves autonomous AI agents out of the cloud and directly onto your local machine. Why it matters: Until now, giving an AI agent control of your operating system…
The first Industrial Revolution was ignited by a small group of engineers, scientists, and industrialists who called themselves the Lunar Society. Figures like James Watt and Josiah Wedgwood did not just invent hardware. They applied the…
High above a busy port, a gigantic container crane swings 40-ton cargo boxes with ease. But when that machine grinds to a halt because of an unexpected motor failure, the ripple effect is immediate. At a major terminal, a single crane…
A fascinating paper published on April 27 offers an ingenious approach to understanding the physical limits of large language models. The author, Bojie Li at PineAI, introduces an incompressibility framework bounded by Shannon entropy to…
In the context of artificial intelligence, we often discuss the human in the loop as a safety mechanism or an auditor. But long before we talked about algorithmic guardrails, I experienced the human loop in a much more profound way.…
When we think of heavy machinery, we usually picture massive excavators in open mines or tractors in open fields. But some of the most critical equipment operates entirely indoors. Consider a mega distribution center. Dozens of forklifts…
The gap between a standard language model summarizing web pages and a system capable of enterprise-grade due diligence is massive. Google DeepMind recently announced a significant architectural step to close that gap with Deep Research…
As James Leffew and I visited Caterpillar Inc. Lafayette to share our work with the facility, Sadia Khalil, PhD (Lead Data Scientist) and Gizem Agar, Ph.D. (Technical Product Manager) on my team visited Sanford, and shared the updates on…
It was a great experience sitting down with Nick Melnychuk on the In Production Podcast. He summarized the core of our conversation perfectly in this post. The industrial sector is rushing to adopt generative AI, but we are consistently…
A towering telehandler on a massive construction site abruptly halts operations. The culprit isn't a blown engine or a catastrophic hydraulic failure. Instead, a few high-tension bolts on its jib have slowly worked themselves loose. It is…
The parameter race gets the headlines, but the AI efficiency race is where production bottlenecks are actually solved. For long context language models, the primary bottleneck is the key-value cache. As context windows expand, storing the…
Artificial intelligence often feels abstract when discussed in a conference room or a laboratory. The true test of these systems occurs when they hit the floor of a high-precision facility like the Caterpillar Inc. Lafayette Facility. I…
A large construction project operates like a massive symphony. Excavators, haul trucks, and cranes are the instruments, and they all need to play in perfect time. However, reality often looks more like chaos. A loader sits idle because the…
When an AI agent fails in production - forgetting previous steps, failing to execute tool calls, or cluttering its context window - engineers often blame the foundational model. However, the issue is rarely the model itself; it is the…
Most organizations are finding themselves pilot-rich but transformation-poor. They have hundreds of active AI deployments, yet these often remain isolated islands of productivity. A recent Harvard Business Review piece by Karim Lakhani at…
In heavy industries from mining to commercial shipping, fuel is liquid gold. Yet, on almost every large-scale construction site or extraction pit, a significant portion of that gold literally goes up in smoke. Consider a busy earthmoving…
For years, the standard approach to building multimodal AI involved stitching disparate systems together. You would have a speech-to-text transcriber, a text-based language model for reasoning, and a text-to-speech generator for the…
The primary challenge in adopting agentic AI is rarely the technology itself. It is the management of the work. While generative capabilities continue to expand rapidly, a massive gap remains on the human side of the equation. Recent data…
We are entering a phase where the user interface itself is generated dynamically. The days of downloading an app and hoping its static menus fit your workflow are ending. I've been experimenting with Google’s Disco, and the implications…
Oil rigs in the ocean, wind farms on distant hills, and surface mines in the desert all share a critical operational constraint. They operate with spotty connectivity and have zero time to wait for cloud computing latency. Consider an oil…
In software engineering, good architecture often rhymes. We are currently seeing a classic machine learning pattern reemerge in the world of autonomous agents: the Generative Adversarial Network. A recent technical deep dive from Anthropic…
The expectation was that artificial intelligence would make jobs easier. Instead, it is increasingly making workers exhausted. A recent Harvard Business Review x Boston Consulting Group (BCG) article highlights a growing phenomenon called…
Servicing a giant piece of equipment can feel like solving a three dimensional puzzle where every single bolt and wire matters. Consider a young field technician facing a daunting task: diagnosing a complex fault in a massive power shovel…
The industry is experiencing a rapid shift from conversational AI to autonomous, long-running agents. Tools like Anthropic's Claude Code, the viral open source project OpenClaw, and NVIDIA's newly announced NeMoClaw are proving that AI can…
For almost a decade, the residual connection has been treated as an untouchable truth in deep learning architecture. Every transformer layer simply adds its new output to a running sum of everything that came before it. A new paper from…
Kicking off my first ever GTC right here in San Jose, and I am thrilled to be on the ground a full day before the keynote. My team members, Navid Khajavi, Sadia Khalil, PhD, and I spent today diving deep into an incredible training…
Deep in the Australian outback, a mechanical ballet plays out in the red dirt. Massive 250-ton haul trucks move iron ore day and night, tracing perfect lines across the terraced pits. The cabs are completely empty. The Problem: The Human…
Building an agentic demo is easy. Shipping that agent safely to production is incredibly hard. After the thrill of a quick prototype, engineering teams usually hit a wall. They spend months building the hidden middleware required for…
When companies talk about Generative AI readiness, the conversation usually focuses on two things: the Data Layer and the Technology Layer. Organizations are spending millions cleaning up their data lakes and securing access to the best…
Imagine a factory maintenance chief actively troubleshooting with a 3D digital replica of a power generator. It sounds like science fiction, but it is rapidly becoming the standard operating procedure for heavy industry. The core problem…
There is an ongoing assumption in AI that to get better multimodal reasoning, you just need to throw more parameters and more tokens at the problem. The industry is currently seeing models trained on over a trillion tokens just to…
For the past few years, artificial intelligence has lived inside a chat window. You type a question, it types an answer, and you do the actual work of moving that information into your systems. That era is ending. The new frontier is the…
We usually have to choose between the "Smart Model" (expensive, slow) and the "Fast Model" (cheap, less capable). Anthropic just broke that curve with Claude Sonnet 4.6. I have been testing it extensively for the past week, and it is…
We are used to AI being a destination. You go to a website to talk to the bot. But the future is likely headless. The agent runs on your infrastructure, and the interface is just wherever you happen to be. I was reading about OpenClaw (the…
Not every company can afford to replace their entire fleet with 2026 models. Most mining and construction sites are running mixed fleets where autonomous haul trucks work alongside bulldozers built thirty years ago. The challenge is that…
The strangest sound on a modern construction site is silence. I recently watched a 30 ton hybrid excavator working in a dense urban zone. When the operator swung the boom, the diesel engine dropped to idle. The machine moved on pure…
We typically assume that as AI agents become more autonomous, the human drifts further away from the controls. We imagine a world where we just click "Go" and walk away for coffee. A new research paper from Anthropic analyzes millions of…
For the last five years, supply chain leaders have had one unspoken job title: Professional Firefighter. They wake up to a disruption, spend the day chasing parts or rescheduling trucks, and go to bed hoping the next day is quieter. They…
I once stood on an oil rig deck in the tropics where the ambient temperature was pushing 110 degrees. The massive diesel generators were running at peak load to keep the operation powered. Suddenly, the high temp alarms started screaming.…
We often build search systems based on lexical matching. Does the query string match the database field? But Airbnb just released a paper detailing why this fails at scale, especially with features like "flexible dates." When the search…
Tax planning is one of the highest value services a financial advisor can offer. It is also the most operationally painful. Usually, it involves requesting PDF tax returns, manually keying data into planning software, and cross referencing…
A veteran mechanic once told me that before a transmission fails, it starts to sing. It is a high-pitched, ghostly whine that sits just above the roar of the diesel engine. To the operator, it sounded normal. But deep inside the planetary…
We are leaving the era where we dump every raw log into the training bucket and hope the model figures it out. Training on petabytes of redundant data is not just expensive. It is inefficient. The new frontier is Data Pruning. How do we…
For years, the consensus has been that accounting would be the first profession to fall to AI. It involves rules, structure, and numbers—the perfect playground for automation. But new field evidence suggests the opposite is happening. I…
We accept that adding parameters makes an LLM smarter. We call these the Scaling Laws. But for Agent Systems, we have been operating on intuition. We add a "manager" agent or split a task into three sub-agents, hoping it works better. It…
The sound of a mining tire blowing out is not a pop. It is a cannon blast. When a 14 foot tall tire ruptures, the explosion doesn't just destroy the rubber; it shatters the rim and grounds a critical production asset for shifts. That…
NVIDIA Omniverse just released the full footage of my presentation from their booth at CES 2026. While our leaders Joe Creed, Ogi Redzic, and Jaime Mineart laid out the strategic roadmap on the keynote stage, I had the privilege of…
In excavators and dozers, hydraulic systems are the arteries that lift, dig, and push. A single compromised seal can escalate into a major failure, bringing a multimillion-dollar operation to a halt. The Data: A giant excavator that…
We are used to spending 80% of our engineering time building the "Happy Path. These are the standard, mundane workflows. The daily data movements, the standard approvals, the repetitive reporting. We spend weeks architecting these because…
In every factory I visit, the Plant Manager knows the constraints better than anyone else. They know the boiler cannot run past 400 degrees. They know Shift A cannot work back-to-back weekends. They know the storage tank only holds 5,000…
For decades, the best diagnostic tool on a mine site was a master mechanic’s ear. They could stand next to a 4,000-horsepower diesel engine, listen to the idle, and tell you if an injector was fouling or a valve was loose. But in modern…
A few weeks ago, I shared a map showing where AI might actually impact the Indian workforce. That visual went viral, but it was just the output. Today, I am releasing the engine behind it. We often see AI exposure studies for the US and…
A new paper from Amazon Web Services (AWS) completely dismantles the assumption that you need a massive model to build effective agents. They took a 350 million parameter model, the OPT-350M from Meta, and fine-tuned it specifically for…
This blog is a fantastic deep dive into how multimodal Generative AI can transform one of the toughest gaps in the maintenance lifecycle - diagnosing root causes after an alarm triggers. The level of detail in the solution design, from…
We talk a lot about "Model Hallucinations." We don't talk enough about "Management Hallucinations" - specifically, the belief that if we just deploy a smart enough tool, employees will happily use it. New data from Harvard Business Review…
If you own a dozer or an excavator, you likely know the rule. Half of your maintenance cost is in the undercarriage. It is the brutal interface between the machine and the earth. Whether it is the abrasive silica of a mine site, the impact…
We spend a lot of time discussing embedding models and RAG accuracy. But what happens when your dataset hits 1.5 billion items and your latency budget is under 100ms? That is the reality for Uber. They recently detailed their migration…
I’m currently at CES2026 in Las Vegas, surrounded by the latest robots and gadgets. But the most significant announcement didn't come from a hardware booth - it came from McKinsey & Company's Global Managing Partner, Bob Sternfels. He just…
Hydraulic energy hides in hoses, accumulators, and valve blocks—until it doesn’t. We often fear the moving parts of a machine. But in heavy equipment, the invisible energy stored at 3,000+ PSI is just as deadly. Pinhole leaks can inject…
Stop trying to "prompt" your way to reliable AI agents. In heavy industry (and enterprise), we don't rely on "magic." We rely on Standard Operating Procedures (SOPs). Yet, when we build AI agents, we often expect a system prompt to replace…
Just finished reading “Performance Hints” by Jeff Dean and Sanjay Ghemawat over the holiday break, and I’m absolutely inspired. Having had the privilege to work at Google and sit in meetings where Jeff would discuss strategies and the…
For the past 2 months, I’ve been digging into a question: Who in India is actually at risk from AI? We usually assume the answer is "software engineers in Bangalore" or "data scientists in Hyderabad. I built the India Iceberg Index -…
Kudos to Manuel Franco de la Peña, Ángel Luis Perales Gómez, and Lorenzo Fernández Maimó (Computer Science Faculty, University of Murcia (FIUM)) for a thoughtful piece of work on explainable AI for time‐series models. They highlight an…
DeepSeek just released their Math-V2 paper, and it represents a fundamental shift in how we train AI for mathematical reasoning. The conventional approach trains models by rewarding correct final answers. This worked well enough to…
Over the Thanksgiving break, I took a deep dive into NVIDIA's full-stack software ecosystem. The strategic case is compelling. I have summarized this in a deck using Google / Google DeepMind's NotebookLM. The promise: 50-500x workflow…
Excited to see the release of PAN—a general, interactable, and long-horizon world model that marks a major leap beyond what large language models (LLMs) alone can achieve. While LLMs have transformed how we interact with information, the…
For optimization enthusiasts like me, this is the path forward: neural networks as dynamic, multi-level optimization systems. Google Research's new Nested Learning paper reframes everything: backpropagation is an associative memory.…
It’s been incredible leading the Future of Manufacturing Digital Platform POC efforts at Caterpillar Inc., pushing the boundaries at the forefront of manufacturing digital platforms using NVIDIA tools. I’m truly grateful to have had a…
For me, staying hands-on—even if only 5-10% of the time—does more than keep my technical skills sharp. It develops real empathy for the challenges my team faces, especially when it comes to the “last 20%” of polish that transforms a good…
As we continue to explore the expanding impact of generative AI in the enterprise, I’ve been thinking deeply about how we move beyond traditional prompt engineering into something far more powerful: context engineering. In real-world…
Earlier this week, I had the privilege of participating in a deeply engaging fireside chat at Caterpillar Inc.’s Analytics Center of Excellence in the Financial Services Division. I was honored to share the stage with Piyush Gundecha, a…
As we continue to explore the expanding capabilities of large language models (LLMs), this recent paper from Apple offers a timely and sobering perspective on where we really stand in the journey toward Artificial General Intelligence…
