Markov, the computer-use data startup founded by Dev Mandal (@DevvMandal) and Harish Ashok (@habril27), released a collection of human-performed CAD workflows on August 3rd for training and evaluating AI agents that operate engineering software through a graphical interface.
https://x.com/DevvMandal/status/2084318299765281183 Mandal said in a two-post thread on X that Markov was publishing 50 tasks across AutoCAD, SOLIDWORKS and Siemens NX. The repository on Hugging Face currently lists 51 workflows spanning eight software groups: AutoCAD, SketchUp, SOLIDWORKS, Revit Architecture, Revit MEP, Siemens NX, Revit Structure and STAAD.Pro.
The broader repository matters because it shows what Mandal and Ashok are trying to sell AI developers: complete, reproducible engineering assignments rather than disconnected screenshots or cursor labels. Each workflow includes a structured prompt, an overview document, input and reference files, the completed CAD deliverable, evaluation rubrics and a screen recording of a human carrying out the task. Markov also captured synchronized mouse and keyboard events, frame timing, narration and a derived trajectory format for computer-use models.
Mandal, Markov's CEO, studied aerospace engineering at IIT Madras and previously worked at Sarvam AI, according to Markov's Y Combinator profile. His earlier projects included doubt.ai, an open-source exam question solver that he said handled more than 5,000 questions from over 700 users. Ashok, who says he is 17 and based in Bangalore, previously built Zenith, a voice assistant for robotics engineers that generated, compiled, uploaded and debugged firmware.
Those backgrounds put both founders close to the problem Markov is pursuing. Engineering programs require long, stateful sequences of clicks, keyboard input, file operations and geometric decisions. A usable agent must preserve context across that sequence and produce a file that can be inspected for dimensions, structure and completeness.
The task count masks uneven workflow lengths
The release totals 99.03 hours and 152 GB. Its 51-workflow headline does not represent equal coverage across the applications. SOLIDWORKS has 19 workflows totaling 16.64 hours, while AutoCAD has four workflows totaling 25.7 hours. SketchUp contributes seven workflows and 25.15 hours. Revit MEP has a single workflow lasting 7.19 hours.
That distribution makes the archive closer to a set of deep demonstrations than a balanced benchmark across CAD products. It can expose a model to complete engineering sessions, but comparisons between applications would need to account for the large differences in task count and recording time.
Markov's release also includes the artifacts needed to score an agent's work. The gold output and rubric can support evaluation against the final deliverable, while the human recording and interaction trace provide a reference path through the software. An agent does not necessarily need to copy that path, provided its final CAD file satisfies the task requirements.
Markov is moving from raw recordings to testable environments
The CAD release extends Markov's earlier work collecting broad computer-use demonstrations. Its Computer Use Large dataset, released in March, contains 48,478 videos representing about 12,300 hours of work in AutoCAD, Blender, Excel, Photoshop, Salesforce and VS Code. Those videos were sourced from the internet, trimmed to remove non-screen content and stripped of audio.
CAD Environments is smaller and more structured. Human operators begin with defined inputs and produce known outputs under an explicit rubric. That makes the new release useful for supervised training, trajectory analysis and final-artifact evaluation, rather than exposure to software use alone.
Markov also published AutoCAD Bench, a separate 50-task benchmark for testing whether multimodal agents can recreate dimensioned 2D drawings and accurate 3D models. The benchmark runs agents inside isolated AutoCAD sessions and evaluates the resulting DWG files. Together, the benchmark and the new workflow archive give Markov both sides of an agent-development pipeline: demonstrations for training and controlled tasks for measurement.
Y Combinator lists Markov as a two-person San Francisco company in its Summer 2026 batch. YC's standard deal invests $500,000 in each accepted company, split between $125,000 for 7% and $375,000 on an uncapped safe with a most-favored-nation provision.
Publishing the CAD archive gives Mandal and Ashok a public demonstration of the collection and packaging system behind Markov. The commercial opportunity sits with AI labs that need proprietary tasks, human demonstrations and evaluation environments for specialized desktop software. CAD is a useful proving ground because an agent's result can be judged through the file it produces, rather than whether its sequence of clicks merely looked plausible.