At the end of some workdays, Andriy Mulyar helps his father review construction contracts. His father, a small-business contractor, works with a notepad and a phone app. His son runs an artificial-intelligence company in New York. Between them lies the familiar comedy of modern progress: astonishing machines on one side, stubborn paperwork on the other, and a human still squinting at the clause that matters.
The scene is small enough to miss. It also contains much of Mulyar's current thesis. The physical world is designed and built through documents that are dense, visual and cross-referenced. A single project can hinge on hundreds of drawing sheets, specifications, submittals and contracts. General-purpose AI can speak fluently about almost anything. Ask it to trace a callout across a construction set and its confidence may survive longer than its context.
Nomic, the company Mulyar co-founded in 2022 and now leads as CEO, is trying to close that gap. Its agents work across architecture, engineering and construction data, where meaning lives not only in words but also in geometry, layout and the relationship between one page and another. It is a narrower mission than Nomic began with. Mulyar prefers a more precise verb: the company did not pivot; it sharpened.
We didn't pivot. We sharpened our focus. We stopped trying to boil the entire ocean and found one sea and focused on it like crazy.Andriy Mulyar
Before the map, the medical chart
Mulyar was born in Ukraine and raised in Virginia, about an hour south of Washington, D.C. He was the first generation of his family in the United States. His public biography starts not with a company but with an appetite. On his personal site, he writes that teaching computers to do things once reserved for people has occupied his mind since about 2017. He also lists the conversational traps most likely to distract him: self-supervised learning, pre-transformer natural-language processing, clinical informatics, data cleaning and Ukraine.
At Virginia Commonwealth University, where he studied mathematics and computer science, that appetite became research. A 2018 fellowship examined decision trees and difficult, imbalanced datasets. He won first place at the college's undergraduate research symposium. The following years took him into clinical language: a summer at Johns Hopkins' Center for Language and Speech Processing, work on medical text tools, and papers on classifying clinical notes and linking concepts buried in them.
A map of the mapmaker
In 2020 he joined Rad AI as an early machine-learning engineer. There, he trained large language models on radiology reports before conversational AI became a dinner-table phrase. The job joined two difficult things: models at scale and language whose errors have consequences. He also met Brandon Duderstadt, who would become his Nomic co-founder.
Mulyar began a machine-learning Ph.D. at NYU Courant in 2021, studying deep-network interpretability. He left in April 2022 to build Nomic. The academic question followed him out the door: how can people see what a model has learned, inspect the material beneath it and correct what went wrong?
The weekend that put AI on a laptop
Nomic's first language for this problem was cartography. Atlas turned huge collections of unstructured information into interactive maps. Instead of querying rows in a table, users could look across clusters of text, images or other data and find patterns, strange neighborhoods and bad examples. Visualization was not decoration. It was a way to make a model's latent space available for inspection.
Then, in March 2023, Nomic released GPT4All. The first model was trained from a curated set of assistant-style examples, and it could run locally on consumer hardware. Mulyar credited Yuvanesh Anand and Zach Nussbaum for an intense weekend of work behind the launch. Within two weeks, the project collected 25,000 GitHub stars. It grew into an ecosystem of local models, bindings and a desktop chat interface.
The zero deserves its place beside the larger figures. GPT4All reached roughly a quarter-million monthly users, but the open project itself generated no revenue. It proved that Nomic could train, compress and distribute models. It created a community. It did not settle the ordinary company question of who would pay for what.
The team kept publishing. Nomic Embed opened weights, training code and data for a long-context text model. A companion vision model placed text and images in a shared latent space. The work carried Mulyar's earlier themes forward: useful systems should be inspectable, reproducible and available to more than a small circle of labs.
A visualization isn't the answer. But an answer without visualization is useless.Andriy Mulyar
Choosing the difficult document
By late 2024, Mulyar was running two questions in parallel. Where did agents fail to create value? And which large industry suffered most from that failure? His answer was long-horizon, multimodal work in the built environment. Construction data is a hostile little kingdom for generic AI: drawings with tiny annotations, references that leap across sheets, specifications that qualify what the picture seems to say, and project histories scattered among systems.
He did not pretend to be an engineer. He bought a ticket to an industry conference in Texas, cold-emailed attendees and arranged three days of meetings with engineering leaders. He described where Nomic stood and asked firms to work as design partners. Several became customers. Then he hired a structural engineer to sit with him daily, translating terminology, workflows and the unspoken rules by which a tool earns trust.
There is an attractive bluntness to the approach. The outsider advantage is usually advertised as freedom from convention. Mulyar paired it with a tax: the outsider has to listen more carefully. In engineering, he concluded, candour earns trust. A clever demo gets a meeting; accurate boundaries keep one.
Engineering is an industry of trust. When you have candour, people trust you. That's what I learned early on.Andriy Mulyar
The harness around the intelligence
Nomic's built-world platform treats drawing review, submittal review, code compliance and project search as variations on a common agent system. The model is one ingredient. Around it sits the harness: document parsing, retrieval, integrations, permissions, cost controls and the machinery that lets an agent keep working across a long task.
AEC-Bench, released by Nomic in April 2026, makes the argument measurable. The open evaluation includes 196 tasks across nine families and three scales, from understanding one sheet to coordinating information across an entire project. Its central finding is refreshingly unromantic. Agents often fail before the glamorous reasoning begins. They cannot reliably retrieve the correct sheet, detail or document.
That is why Nomic trains domain-specific parsing and embedding models. A flat text extractor can turn a drawing into a bag of labels while discarding the lines that tell those labels where they belong. A vision model can spend vast numbers of tokens looking at a high-resolution sheet and still miss a precise relationship. The system needs structured representations and tools made for the document, not only a more eloquent model.
Mulyar's product boundary is equally concrete. An agent should begin where the human begins work and end where the human ends. If its result must be copied through three extra systems and translated by an enthusiast in the corner, daily adoption will wither. Nomic connects to the project repositories and delivery software where drawings already live, then returns markups and findings into the workflow.
The other eighty percent
Technical founders enjoy the portion of a problem that compiles. Mulyar assigns only 20 percent of the challenge to building a working product. The remaining 80 percent is education and change management: explaining what an agent is, teaching people how to verify its output, controlling cost and helping a large firm deploy it without producing expensive nonsense.
It is a less cinematic picture of AI transformation. The future arrives with training sessions, contract reviews, jurisdictional requirements and an employee asking whether the result can be trusted. Mulyar has spent stretches personally training the people who train customers. The company has hired architects and engineers alongside machine-learning researchers. Domain expertise is treated as operating infrastructure.
The ambition remains large even after the narrowing. Nomic wants agents to help the people who design, construct and maintain physical assets move through institutional knowledge with fewer delays and less rework. Mulyar's monthly “AI in the Built World” commentary follows the market as models enter homebuilding, engineering and construction operations. His public speaking has shifted accordingly, from how to build GPT4All to how an agent reviews a multi-sheet drawing set.
There is continuity beneath the change. The undergraduate studying interpretable decision trees, the clinical-NLP engineer structuring difficult notes, the founder mapping embedding spaces and the CEO parsing blueprints are all pursuing the same stubborn idea: intelligence becomes useful when people can locate the evidence, inspect the path and correct the error.
Late in the day, that idea returns to the contractor, the notepad and the phone. The contract is not a benchmark. It is work waiting to be understood. Mulyar's bet is that AI will matter in the built world when it can meet that document where it lives, follow its references without losing the plot and hand a person something worth trusting. The blueprint, after all, was always a map.