We got a lot of heat for promoting Research as a Service. Questions were asked: Can we just buy papers? Is thus a paper mill? Unfortunately, what most people don't know is that this has existed as a business model for decades. Especially in pharma, where drug development is painful and often the R&D is outsourced away to smaller, leaner companies that can iterate faster. We are simply doing this for ML and AI. And it absolutely works! There is one part that deserves particular clarity: you cannot buy a peer-reviewed paper. We cannot and don't guarantee that NeurIPS, CVPR, ICLR, AAAI, or any other venue will accept a piece of work, and nobody credible can. Peer review exists precisely because an independent community decides whether the research meets its standards. Shared authorship makes sense when both parties make substantive research contributions and must follow the authorship and disclosure rules of the relevant venue. And this is exactly what we did and our Research-as-a-Service collab with Renesas went in NeurIPS 2026. Both names on the paper, like the card says. We established CausalDriveBench, which tests whether modern vision-language-action models used for autonomous driving actually reason about cause and effect, rather than simply producing plausible explanations alongside plausible driving trajectories.

We are terribly impatient as a company and absolutely hate mulling over things for six months before they exist. We would rather start building the thing, and course-correct on the way. If you want to automate part of your physical factory, call us. If you are a fast-growing startup with a difficult research problem but cannot justify building an entire research team, call us. If your company has decades of knowledge scattered across drawings, PDFs, databases and people’s heads, and you want to turn it into something your machines can actually understand, call us. Or just talk to @SanjaySuriya6 who is in SF now!

We believe sales should be done by engineers themselves. So we only hire people with a body count of 0 or less. Speaking of sales, @SanjaySuriya6 and @cgnarendiran are in SF to establish our presence there. They've spent the last few years solving hard ML problems for large enterprises (Bosch, Mercedes-Benz) and cutting-edge startups (Smallest AI, Entelligence) If you're a founder working on problems like: - Post-training - Voice AI (Cascaded, S2S, Full-duplex) - Physical AI (World models, VLAs, ADAS) They would love to grab coffee and explore how they can help you tackle your technical challenges. Currently, pipeline is the only body count we’re currently trying to increase.

@SanjaySuriya6 was a physics student who wanted to get into ML. He went through every professor at IISc working on deep learning and wrote to all of them. @Arjunjain was one. The video tells you how that went. He's now a partner at FastCode AI, a research scientist who also runs our forward-deployed engineering, the team that stays with a client until the model works in production. He's in San Francisco with Naren and @siddesh_239. It's the same cold email, just on camera. P.S. - Sent from the FastCode AI account or he'd have undersold himself.

Last month, one of our customers nearly shut down a plant. For a large process plant, even a single day of unplanned downtime can easily run into a few million dollars. Yes, you read that right, for one single day. The problem was not some exotic failure. The information the engineers needed already existed, but it was buried across engineering drawings that had to be read, traced, and reconciled manually. Work that would normally take three days was completed in roughly three hours. That incident captures why we built Legend. Legend is our framework for turning engineering drawings into one machine-readable model of an industrial operation. Once the drawings are understood as a connected system rather than a pile of files, teams can start asking much more useful questions across the plant. How many pumps of a certain type are installed? Which equipment needs to be isolated before maintenance? What is connected upstream or downstream of a component? Which parts of the plant are affected if something changes? No engineer can realistically hold thousands of drawings and all of their relationships in their head at once. Software can. That is the larger idea behind Legend. We do not want engineering drawings to remain static documents that become useful only when somebody manually opens them and starts tracing lines. We want them to become part of the operational intelligence of the plant. Legend is already deployed across 40 industrial sites, including oil and gas facilities, and we are now extending the same approach into hydraulics, pneumatics, electrical systems, and automation. And because industrial data is often among the most sensitive data a company owns, Legend can run on-premise and air-gapped, on infrastructure controlled by the customer. Every drawing has a legend. We just built one for the physical world.

Reading a drawing is one thing. Understanding the system behind it is another. Today we’re introducing Legend™, Fast Code AI’s Engineering Intelligence system. Engineering drawings contain far more than symbols, tags and lines. They describe how equipment connects, how systems behave, where dependencies exist and how an operation was designed to work. Most of that knowledge is still locked inside drawings that were created for engineers to read, not machines. Legend turns those drawings into connected, machine-readable engineering context. It identifies the elements on the page, reconstructs their relationships across drawings, and creates a structured representation of the system behind them. We started working on this problem in 2019. Today, the technology is running in an industrial environment, and we are extending it across engineering domains. Every drawing has a legend. We built one for your entire operation.

Apple shipped a foldable seven years after everyone else. It waited until it was really good, not first. There is a lesson in that for anyone buying AI. The vendor who demos first is rarely the one still standing last. First is a weekend. Best is the version that holds up when the drawing is smudged, the audit is real, and one wrong reading shuts the line down. We waited years to ship 𝗟𝗲𝗴𝗲𝗻𝗱. It reads the drawings that run plants, and it runs at a major energy operator today. Buy the one that works, not the one that was early.

AI models discovered engineering drawings in 2023. This chart is from a new survey in Artificial Intelligence Review - the shading marks "rapid adoption in engineering." Now read the same survey's fine print. SAM can famously segment anything - except a pipe on a P&ID. It sees in 16×16 pixel patches, a pipe is one pixel wide, and inside the patch the line just gets ignored. CLIP learned to see from captions, and captions say what is in an image, never where. On a drawing, where is everything. DINOv2 - the serious one, the backbone behind today's best inspection results - earned that crown reading texture: cracks, porosity, surface grain. A drawing has no texture, and even DINOv2 drops 12 to 20 points on drawings. GPT-4o scores 74.9% on the industrial inspection benchmark. Production needs better than 99. And the near-perfect scores you do see live on lab benchmarks that the survey itself calls saturated - they fall apart in the real world. Foundation models are amazing. They still cannot read an engineering drawing. We learned this the hard way, because we started in 2019 - before this chart begins, before the tools existed. Nothing off the shelf worked. And the models were not even the hardest part. The hardest part was that we had to learn to read these drawings ourselves first - flowlines, isolations, spec breaks, the HMB table, the EPS sections, equipment, valves, instruments, electrical. Engineers spend years learning this language. We fell a little in love with it. There is a strange reward in the moment a sheet that looked like noise starts reading like the story of a plant. So we built Legend™: one system, one job, reading the drawings that run plants. It took seven years. It runs at a major energy operator today, and hydraulics and HVAC companies are now at the table. Bring us your hardest drawing. (Figure and every number above: Kishore Reddy Pagidi & Anusha Gardas, Artificial Intelligence Review, 2026) Activate to view larger image,

In climbing, they don't call them routes. They call them problems. The team spent this morning on the wall - bouldering and top rope. But what got me was the competitiveness. Nobody was willing to give up. Least of all on the dyno: the move where you leap off the wall, both hands flying at the same hold, and either you stick it and adjust or you drop to the mat. We threw ourselves at it again and again, laughing, falling, chalking up, going again - refusing to walk away until they stuck it. FTR Mitish, Siddharth and Harsh stuck it. It is oddly close to how we spend our days, just with chalk instead of a keyboard. Find a hard problem. Fail on it in the open. Throw yourself at it again. Send it. Belay each other on the way up. Soccer, then pickleball, now climbing. Turns out the people who won't quit a hard problem at work, won't quit one on a wall either.

We are a humble AI lab building models that can actually read engineering drawings, starting with Piping and Instrumentation Diagrams (P&IDs). The reason we exist is because frontier AI, even with its Nobel-Laureate level intelligence, can't parse standard engineering diagrams. Before AI decides to go on this killing spree and take over the world, we would humbly request that it first learn to trace a pipe through a slightly blurry P&ID without hallucinating a valve. Until then, we remain cautiously optimistic about humanity’s chances.