Buildings 2.0 Library
Conversations on the future of the built environment
81 episodes of the Buildings 2.0 podcast, each bound as a short book: the ideas, the numbers and the lines worth keeping.
Season 1, Episode 43 /Jan 16, 2025
Democratizing tech tools for AEC
Mayur Mistry, 3D Guru
Jan 16, 2025 /AI & Automation
Democratizing tech tools for AEC
Mayur Mistry co-founded 3DGuru.ai, a startup building AI tools for architecture and construction, after training as a structural engineer in India and working in Silicon Valley. He talks about a shift he has watched firsthand: the cost of training a custom AI model collapsing from a specialist's project to the price of a coffee, and what that means for firms sitting on decades of their own drawings.
Mayur Mistry, Co-Founder, 3D Guru
In this conversation
Mayur's figure for what a firm-specific model costs now versus what it cost seven years ago, and why that changes who gets to build AI.
Why an internal tools team, a full-stack engineer, a product manager and a machine-learning engineer, can cost a quarter million dollars a year with few examples of it paying off.
The prerequisite most firms skip: a data strategy that aggregates structured and unstructured project data so a model can read what matters inside a 3D model.
His advice to start from the real-world catalog of materials and finishes before adding AI, so the output can actually be bought and built.
Now I can train a model using five images and it will cost me less than a Starbucks coffee.
Mayur Mistry
Chapters
Five images and a cup of coffee
How far the cost of a custom AI model has fallen, and what a firm should do about it.
By the numbers
Fall in the cost of running a GPT-3.5-level model, November 2022 to October 2024
Mayur says a firm-specific model now costs less than a coffee. Stanford's AI Index tracks the same collapse in price at the industry level.
Source: Stanford HAI, AI Index Report 2025 ↗From Bangalore to Silicon Valley
A structural engineer's path from India to a Silicon Valley startup.
Everyone can build their own Alexa
Why voice assistants and chat models changed who can build software.
Trial, error, and repeat on steroids
What it took to learn the startup game by repetition.
Charge for pilots sooner
The go-to-market lessons Mayur would apply sooner if he started again.
You always trade general intelligence versus specialized intelligence.
Mayur Mistry
Takeaways
Training a generative model needed thousands of images, a GPU and a specialist in 2018. Today a firm can train one on five images for a few dollars, which makes a firm's own data, not the model, the asset.
General intelligence versus specialized intelligence is the real trade. A firm that trains on its own details and style can beat a general model at its own work.
Charge for pilots. Free pilots produce polite feedback and no commitment; paid ones produce the truth and a customer.
A lot of the AI tools now are 80% there. We need that domain specific knowledge to make it really useful.
Mayur Mistry